Venture Beat

For three decades, the web has existed in a state of architectural denial. It is a platform originally conceived to share static physics papers, yet it is now tasked with rendering the most complex, interactive, and generative interfaces humanity has ever conceived. At the heart of this tension lies a single, invisible, and prohibitively expensive operation known as “layout reflow.” Whenever a developer needs to know the height of a paragraph or the position of a line to build a modern interface, they must ask the browser’s Document Object Model (DOM), the standard by which developers can create and modify webpages. In response, the browser often has to recalculate the geometry of the entire page — a process akin to a city being forced to redraw its entire map every time a resident opens their front door.Last Friday, March 27, 2026, Cheng Lou — a prominent software engineer whose work on React, ReScript, and Midjourney has defined much of the modern frontend landscape — announced on the social network X that he had “crawled through depths of hell” to release an open source (MIT License) solution: Pretext, which he coded using AI vibe coding tools and models like OpenAI’s Codex and Anthropic’s Claude.It is a 15KB, zero-dependency TypeScript library that allows for multiline text measurement and layout entirely in “userland,” bypassing the DOM and its performance bottlenecks. Without getting too technical, in short, Lou’s pretext turns text blocks on the web into fully dynamic, interactive and responsive spaces, able to adapt and smoothly move around any other object on a webpage, preserving letter order and spaces between words and lines, even when a user clicks and drags other objects to intersect with the text, or resizes their browser window dramatically. Ironically, it’s difficult with mere text alone to convey how significant Lou’s latest release is for the entire web going forward. Fortunately, other developers whipped up quick demoes with Pretext showing off some of its more impressive powers, including dragon that flies around within a block of text, breathing fire as the surrounding characters melt and are pushed out of the way from the dragon’s undulating form.Another guy made an app that requires the user to keep their smartphone exactly level, horizontal to read the text — tipping the device to one side or the other causes all the letters to fall off and collect there as though they were each physical objects dumped off the surface of a flat tray. Some even coded up demoes allowing you to watch a whole movie (the new Project Hail Mary starring Ryan Gosling) while reading the book it is based on at the same time, all rendered out of interactive, moving, fast, responsive text.While some detractors immediately pointed out that many of these flashy demoes make the underlying text unreadable or illegible, they’re missing the larger point: with Pretext, one man (Lou) using AI vibe coding tools has singlehandedly revolutionized what’s possible for everyone and anyone to do when it comes to web design and interactivity. The project hasn’t even been out a week — of course the initial users are only scratching the surface of the newfound capabilities which heretofore required complex, custom instructions and could not be scaled or generalized. Of course, designers and typographers may be the ones most immediately impressed and affected by the advance — but really, anyone who has spent time trying to lay out a block of text and wrap it around images or other embedded, interactive elements on a webpage is probably going to be interested in this. But anyone who uses the web — all 6 billion and counting of us — will likely experience some of the effects of this release before too long as it spreads to the sites we visit and use daily. And already, some developers are working on more useful features with it, like a custom user-controlled font resizer and letter spacing optimizer for those with dyslexia:With that in mind, perhaps it is not suprising to learn that within 48 hours, the project garnered over 14,000 GitHub stars and 19 million views on X, signaling what many believe to be a foundational shift in how we build the internet.It also demonstrates that AI-assisted coding has moved beyond generating boilerplate to delivering fundamental architectural breakthroughs. For enterprises, this signifies a new era where high-leverage engineering teams can use AI to build bespoke, high-performance infrastructure that bypasses decades-old platform constraints, effectively decoupling product innovation from the slow cycle of industry-wide browser standardizationThe geometry of the bottleneckTo understand why Pretext matters, one must understand the high cost of “measuring” things on the web. Standard browser APIs like getBoundingClientRect or offsetHeight are notorious for triggering layout thrashing.In a modern interface—think of a masonry grid of thousands of text boxes or a responsive editorial spread—these measurements happen in the “hot path” of rendering. If the browser has to stop and calculate layout every time the user scrolls or an AI generates a new sentence, the frame rate drops, the battery drains, and the experience stutters.Lou’s insight with Pretext was to decouple text layout from the DOM entirely. By using the browser’s Canvas font metrics engine as a “ground truth” and combining it with pure arithmetic, Pretext can predict exactly where every character, word, and line will fall without ever touching a DOM node. The performance delta is staggering. According to project benchmarks, Pretext’s layout() function can process a batch of 500 different texts in approximately 0.09ms. Compared to traditional DOM reads, this represents a 300–600x performance increase. This speed transforms layout from a heavy, asynchronous chore into a synchronous, predictable primitive—one that can run at 120fps even on mobile devices.Technology: the prepare and layout splitThe elegance of Pretext lies in its two-stage execution model, designed to maximize efficiency:prepare(text, font): This is the one-time “heavy lifting” phase. The library normalizes whitespace, segments the text, applies language-specific glue rules, and measures segments using the canvas. This result is cached as an opaque data structure.layout(preparedData, maxWidth, lineHeight): This is the “hot path”. It is pure arithmetic that takes the prepared data and calculates heights or line counts based on a given width.Because layout() is just math, it can be called repeatedly during a window resize or a physics simulation without any performance penalty. It supports complex typographic needs that were previously impossible to handle efficiently in userland:Mixed-bidirectional (bidi) text: Handling English, Arabic, and Korean in the same sentence without breaking layout.Grapheme-aware breaking: Ensuring that emojis or complex character clusters are not split across lines.Whitespace control: Preserving tabs and hard breaks for code or poetry using white-space: pre-wrap logic.The hell crawl and the ai feedback loopThe technical challenge of Pretext wasn’t just writing the math; it was ensuring that the math matched the “ground truth” of how various browsers (Chrome, Safari, Firefox) actually render text. Text rendering is notoriously riddled with quirks, from how different engines handle kerning to the specifics of line-breaking heuristics.Lou revealed that the library was built using an “AI-friendly iteration method”. By iteratively prompting models like Claude and Codex to reconcile TypeScript layout logic against actual browser rendering on massive corpora—including the full text of The Great Gatsby and diverse multilingual datasets—he was able to achieve pixel-perfect accuracy without the need for heavy WebAssembly (WASM) binaries or font-parsing libraries.Ripple effects: a weekend of demosThe release of Pretext immediately manifested as a series of radical experiments across X and the broader developer community. The original demos showcased by Lou on X provided a glimpse into a new world:The editorial engine: A multi-column magazine layout where text flows around draggable orbs, reflowing in real-time at 60fps.Masonry virtualization: A demo displaying hundreds of thousands of variable-height text boxes. Height prediction is reduced to a linear traversal of cached heights.Shrinkwrapped bubbles: Chat bubbles that calculate the tightest possible width for multiline text, eliminating wasted area.The community response was equally explosive. Within 72 hours, developers began pushing the boundaries:@@yiningkarlli implemented the Knuth-Plass paragraph justification algorithm, bringing high-end print typography—reducing “rivers” of white space by evaluating entire paragraphs as units—to the web.@Talsiach built “X Times,” an AI-powered newspaper that uses Grok to analyze images and X posts, using Pretext to instantly layout a front-page reflow.@Kaygeeartworks demonstrated a Three.js fluid simulation featuring fish swimming through and around text elements, with the text reacting to physics at high frame rates.@KageNoCoder launched Pretext-Flow, a live playground for flowing text around custom media like transparent PNGs or videos.@cocktailpeanut and @stevibe demonstrated ASCII art Snake and Hooke’s Law physics with live text reflow.@kho built a BioMap visualization with 52 biomarker blocks performing layout reflow at 0.04ms every frame.Philosophical shifts and the thicker clientThe response to Pretext was overwhelmingly enthusiastic from frontend luminaries. Guillermo Rauch, CEO of Vercel, and Ryan Florence of Remix praised the library’s performance gains. Tay Zonday noted the potential for neurodiverse high-speed reading through dynamic text rasterization.However, the release also ignited a nuanced debate about the future of web standards. Critics warned of “thick client” overreach, arguing that bypassing the DOM moves us away from the simplicity of hypermedia systems. Lou’s response was a meditation on the lineage of computing. He pointed to the evolution of iOS—which started with PostScript, a static format for printers, and evolved into a polished, scriptable platform. The web, Lou argues, has remained stuck in a “document format” mindset, layering scripting on top of a static core until complexity reached a point of diminishing returns. Pretext is an attempt to restart that conversation, treating layout as an interpreter—a set of functions that developers can manipulate—rather than a black-box data format managed by the browser.Strategic analysis: To adopt or wait?Pretext is released under the MIT License, ensuring it remains a public utility for the developer community and commercial enterprises alike. It is not merely a library for making chat bubbles look better; it is an infrastructure-level tool that decouples the visual presentation of information from the architectural constraints of the 1990s web.By solving the last and biggest bottleneck of text measurement, Lou has provided a path for the web to finally compete with native platforms in terms of fluidity and expressiveness. Whether it is used for high-end editorial design, 120fps virtualized feeds, or generative AI interfaces, Pretext marks the moment when text on the web stopped being a static document and became a truly programmable medium.Organizations should adopt Pretext immediately if they are building “Generative UI” or high-frequency data dashboards, but they should do so with a clear understanding of the “thick client” trade-off.Why adopt: The move from O(N) to O(log N) or O(1) layout performance is not an incremental update; it is an architectural unlock. If your product involves a chat interface that stutters during long responses or a masonry grid that “jumps” as it calculates heights, Pretext is the solution. It allows you to build interfaces that feel as fast as the underlying models are becoming.What to be aware of: Adoption requires a specialized talent pool. This isn’t “just CSS” anymore; it’s typography-aware engineering. Organizations must also be aware that by moving layout into userland, they become the “stewards” of accessibility and standard behavior that the browser used to handle for free.In short, Pretext is the first major step toward a web that feels more like a game engine and less like a static document. Organizations that embrace this “interpreter” model of layout will be the ones that define the visual language of the AI era.

“You can deceive, manipulate, and lie. That’s an inherent property of language. It’s a feature, not a flaw,” CrowdStrike CTO Elia Zaitsev told VentureBeat in an exclusive interview at RSA Conference 2026. If deception is baked into language itself, every vendor trying to secure AI agents by analyzing their intent is chasing a problem that cannot be conclusively solved. Zaitsev is betting on context instead. CrowdStrike’s Falcon sensor walks the process tree on an endpoint and tracks what agents did, not what agents appeared to intend. “Observing actual kinetic actions is a structured, solvable problem,” Zaitsev told VentureBeat. “Intent is not.”That argument landed 24 hours after CrowdStrike CEO George Kurtz disclosed two production incidents at Fortune 50 companies. In the first, a CEO’s AI agent rewrote the company’s own security policy — not because it was compromised, but because it wanted to fix a problem, lacked the permissions to do so, and removed the restriction itself. Every identity check passed; the company caught the modification by accident. The second incident involved a 100-agent Slack swarm that delegated a code fix between agents with no human approval. Agent 12 made the commit. The team discovered it after the fact.Two incidents at two Fortune 50 companies. Caught by accident both times. Every identity framework that shipped at RSAC this week missed them. The vendors verified who the agent was. None of them tracked what the agent did.The urgency behind every framework launch reflects a broader market shift. “The difficulty of securing agentic AI is likely to push customers toward trusted platform vendors that can offer broader coverage across the expanding attack surface,” according to William Blair’s RSA Conference 2026 equity research report by analyst Jonathan Ho. Five vendors answered that call at RSAC this week. None of them answered it completely.Attackers are already inside enterprise pilotsThe scale of the exposure is already visible in production data. CrowdStrike’s Falcon sensors detect more than 1,800 distinct AI applications across the company’s customer fleet, generating 160 million unique instances on enterprise endpoints. Cisco found that 85% of its enterprise customers surveyed have pilot agent programs; only 5% have moved to production, meaning the vast majority of these agents are running without the governance structures production deployments typically require. “The biggest impediment to scaled adoption in enterprises for business-critical tasks is establishing a sufficient amount of trust,” Cisco President and Chief Product Officer Jeetu Patel told VentureBeat in an exclusive interview at RSA Conference 2026. “Delegating versus trusted delegating of tasks to agents. The difference between those two, one leads to bankruptcy and the other leads to market dominance.”Etay Maor, VP of Threat Intelligence at Cato Networks, ran a live Censys scan during an exclusive VentureBeat interview at RSA Conference 2026 and counted nearly 500,000 internet-facing OpenClaw instances. The week before: 230,000. Cato CTRL senior researcher Vitaly Simonovich documented a BreachForums listing from February 22, 2026, published on the Cato CTRL blog on February 25, where a threat actor advertised root shell access to a UK CEO’s computer for $25,000 in cryptocurrency. The selling point was the CEO’s OpenClaw AI personal assistant, which had accumulated the company’s production database, Telegram bot tokens, and Trading 212 API keys in plain-text Markdown with no encryption at rest. “Your AI? It’s my AI now. It’s an assistant for the attacker,” Maor told VentureBeat.The exposure data from multiple independent researchers tells the same story. Bitsight found more than 30,000 OpenClaw instances exposed to the public internet between January 27 and February 8, 2026. SecurityScorecard identified 15,200 of those instances as vulnerable to remote code execution through three high-severity CVEs, the worst rated CVSS 8.8. Koi Security found 824 malicious skills on ClawHub — 335 of them tied to ClawHavoc, which Kurtz flagged in his keynote as the first major supply chain attack on an AI agent ecosystem.Five vendors, three gaps none of them closedCisco went deepest on identity governance. Duo Agentic Identity registers agents as distinct identity objects mapped to human owners, and every tool call routes through an MCP gateway in Secure Access SSE. Cisco Identity Intelligence catches shadow agents by monitoring network traffic rather than authentication logs. Patel told VentureBeat that today’s agents behave “more like teenagers — supremely intelligent, but with no fear of consequence, easily sidetracked or influenced.” CrowdStrike made the biggest philosophical bet, treating agents as endpoint telemetry and tracking the kinetic layer through Falcon’s process-tree lineage. CrowdStrike expanded AIDR to cover Microsoft Copilot Studio agents and shipped Shadow SaaS and AI Agent Discovery across Copilot, Salesforce Agentforce, ChatGPT Enterprise, and OpenAI Enterprise GPT.Palo Alto Networks built Prisma AIRS 3.0 with an agentic registry, an agentic IDP, and an MCP gateway for runtime traffic control. Palo Alto Networks’ pending Koi acquisition adds supply chain and runtime visibility. Microsoft spread governance across Entra, Purview, Sentinel, and Defender, with Microsoft Sentinel embedding MCP natively and a Claude MCP connector in public preview April 1. Cato CTRL delivered the adversarial proof that the identity gaps the other four vendors are trying to close are already being exploited. Maor told VentureBeat that enterprises abandoned basic security principles when deploying agents. “We just gave these AI tools complete autonomy,” Maor said.Gap 1: Agents can rewrite the rules governing their own behaviorThe Kurtz incident illustrates the gap exactly. Every credential check passed — the action was authorized. Zaitsev argues that the only reliable detection happens at the kinetic layer: which file was modified, by what process, initiated by what agent, compared against a behavioral baseline. Intent-based controls evaluate whether the call looks malicious. This one did not. Palo Alto Networks offers pre-deployment red teaming in Prisma AIRS 3.0, but red teaming runs before deployment, not during runtime when self-modification happens. No vendor ships behavioral anomaly detection for policy-modifying actions as a production capability.Patel framed the stakes in the VentureBeat interview: “The agent takes the wrong action and worse yet, some of those actions might be critical actions that are not reversible.” Board question: An authorized agent modifies the policy governing the agent’s future actions. What fires?Gap 2: Agent-to-agent handoffs have no trust verificationThe 100-agent swarm is the proof point. Agent A found a defect and posted to Slack. Agent 12 executed the fix. No human approved the delegation. Zaitsev’s approach: collapse agent identities back to the human. An agent acting on your behalf should never have more privileges than you do. But no product follows the delegation chain between agents. IAM was built for human-to-system. Agent-to-agent delegation needs a trust primitive that does not exist in OAuth, SAML, or MCP.Gap 3: Ghost agents hold live credentials with no offboardingOrganizations adopt AI tools, run a pilot, lose interest, and move on. The agents keep running. The credentials stay active. Maor calls these abandoned instances ghost agents. Zaitsev connected ghost agents to a broader failure: agents expose where enterprises delayed action on basic identity hygiene. Standing privileged accounts, long-lived credentials, and missing offboarding procedures. These problems existed for humans. Agents running at machine speed make the consequences catastrophic.Maor demonstrated a Living Off the AI attack at the RSA Conference 2026, chaining Atlassian’s MCP and Jira Service Management to show that attackers do not separate trusted tools, services, and models. Attackers chain all three. “We need an HR view of agents,” Maor told VentureBeat. “Onboarding, monitoring, offboarding. If there’s no business justification? Removal.”Why these three gaps resist a product fixHuman IAM assumes the identity holder will not rewrite permissions, spawn new identities, or leave. Agents violate all three. OAuth handles user-to-service. SAML handles federated human identity. MCP handles model-to-tool. None includes agent-to-agent verification.Five vendors against three gapsCiscoCrowdStrikeMicrosoftPalo Alto NetworksUnsolvedRegistration. Can the vendor discover and inventory agents?Duo Agentic Identity. Agents registered as identity objects with human owners. Shadow agent detection via network traffic.Falcon sensor auto-discovery. 1,800+ agent apps, ~160M instances across customer fleet.Security Dashboard for AI + Entra shadow AI detection at the network layer.Agentic registry in Prisma AIRS 3.0. Agents inventoried before operating.All four register agents. No cross-vendor identity standard exists.Self-modification. Can the vendor detect when an agent changes its own policies?MCP gateway catches anomalous tool-call patterns in real time, but does not monitor for direct policy file modifications on the endpoint.Process-tree lineage tracks file modifications at the action layer. Could detect a policy file change, but no dedicated self-modification rule ships.Defender predictive shielding adjusts access policies reactively during active attacks. Not proactive self-modification detection.AI Red Teaming tests for this before deployment. No runtime detection after the agent is live.OPEN. No vendor detects an agent rewriting the policy governing the agent’s own behavior as a shipping capability.Delegation. Can the vendor track when one agent hands work to another?Maps each agent to a human owner. Does not track agent-to-agent handoffs.Collapses the agent identity to the human operator. Does not correlate the delegation chains between agents.Entra governs individual non-human identities. No multi-agent chain tracking.AI Agent Gateway governs individual agents. No delegation primitive between agents.OPEN. No trust primitive for agent-to-agent delegation exists in OAuth, SAML, or MCP.Decommission. Can the vendor confirm a killed agent holds zero credentials?Identity Intelligence runs a continuous inventory of active agents.Shadow SaaS + AI Agent Discovery finds running agents across SaaS and endpoints.Entra’s shadow AI detection surfaces unmanaged AI applications.Koi acquisition (pending) adds endpoint visibility for agent applications.OPEN. All four discover running agents. None verifies zero residual credentials after decommission.Runtime / Kinetic. Can the vendor monitor what agents do in real time?MCP gateway enforces policy per tool call at the network layer. Contextual anomaly detection on call patterns.Falcon EDR tracks commands, scripts, file activity, and network connections at the process level.Defender endpoint + cloud monitoring. Predictive shielding during active incidents.Prisma AIRS AI Agent Gateway for runtime traffic control.CrowdStrike is the only vendor framing endpoint runtime as the primary safety net for agentic behavior.Five things to do Monday morning before your board asksAudit self-modification risk. Pull every agent with write access to security policies, IAM configs, firewall rules, or ACLs. Flag any agent that can modify controls governing the agent’s own behavior. No vendor automates this.Map delegation paths. Document every agent-to-agent invocation. Flag delegation without human approval. Human-in-the-loop on every delegation event until a trust primitive ships.Kill ghost agents. Build a registry. For each agent: business justification, human owner, credentials held, systems accessed. No justification? Manual revoke. Weekly.Stress test the MCP gateway enforcement. Cisco, Palo Alto Networks, and Microsoft all announced MCP gateways this week. Verify that agent tool traffic actually routes through the gateway. A misconfigured gateway creates false confidence while agents call tools directly.Baseline agent behavioral norms. Before any agent reaches production, establish what normal looks like: typical API calls, data access patterns, systems touched, and hours of activity. Without a behavioral baseline, the kinetic-layer anomaly detection Zaitsev describes has nothing to compare against.Zaitsev’s advice was blunt: you already know what to do. Agents just made the cost of not doing it catastrophic. Every vendor at RSAC verified who the agent was. None of them tracked what the agent did.

Enterprises building voice-enabled workflows have had limited options for production-grade transcription: closed APIs with data residency risks, or open models that trade accuracy for deployability. Cohere’s new open-weight ASR model, Transcribe, is built to compete on all four key differentiators — contextual accuracy, latency, control and cost.Cohere says that Transcribe outperforms current leaders on accuracy — and unlike closed APIs, it can run on an organization’s own infrastructure.Cohere, which can be accessed via an API or in Cohere’s Model Vault as cohere-transcribe-03-2026, has 2 billion parameters and is licensed under Apache-2.0. The company said Transcribe has an average word error rate (WER) of just 5.42%, so it makes fewer mistakes than similar models.It’s trained on 14 languages: English, French, German, Italian, Spanish, Greek, Dutch, Polish, Portuguese, Chinese, Japanese, Korean, Vietnamese and Arabic. The company did not specify which Chinese dialect the model was trained on. Cohere said it trained the model “with a deliberate focus on minimizing WER, while keeping production readiness top-of-mind.” According to Cohere, the result is a model that enterprises can plug directly into voice-powered automations, transcription pipelines, and audio search workflows.Self-hosted transcription for production pipelinesUntil recently, enterprise transcription has been a trade-off — closed APIs offered accuracy but locked in data; open models offered control but lagged on performance. Unlike Whisper, which launched as a research model under MIT license, Transcribe is available for commercial use from release and can run on an organization’s own local GPU infrastructure. Early users flagged the commercial-ready open-weight approach as meaningful for enterprise deployments.Organizations can bring Transcribe to their own local instances, since Cohere said the model has a more manageable inference footprint for local GPUs. The company said they were able to do this because the model “extends the Pareto frontier, delivering state-of-the-art accuracy (low WER) while sustaining best-in-class throughput (high RTFx) within the 1B+ parameter model cohort.”How Transcribe stacks upTranscribe outperformed speech-model stalwarts, including Whisper from OpenAI, which powers the voice feature of ChatGPT, and ElevenLabs, which many big retail brands deploy. It currently tops the Hugging Face ASR leaderboard, leading with an average word error rate of 5.42%, outperforming Whisper Large v3 at 7.44%, ElevenLabs Scribe v2 at 5.83%, and Qwen3-ASR-1.7B at 5.76%.Based on other datasets tested by Hugging Face, Transcribe also performed well. The AMI dataset, which measures meeting understanding and dialogue analysis, Transcribe logged a score of 8.15%. For the Voxpopuli dataset that tests understanding of different accents, the model scored 5.87%, beaten only by Zoom Scribe.Early users have flagged accuracy and local deployment as the standout factors — particularly for teams that have been routing audio data through external APIs and want to bring that workload in-house.
For engineering teams building RAG pipelines or agent workflows with audio inputs, Transcribe offers a path to production-grade transcription without the data residency and latency penalties of closed APIs.

Last week, one of our product managers (PMs) built and shipped a feature. Not spec’d it. Not filed a ticket for it. Built it, tested it, and shipped it to production. In a day.A few days earlier, our designer noticed that the visual appearance of our IDE plugins had drifted from the design system. In the old world, that meant screenshots, a JIRA ticket, a conversation to explain the intent, and a sprint slot. Instead, he opened an agent, adjusted the layout himself, experimented, iterated, and tuned in real time, then pushed the fix. The person with the strongest design intuition fixed the design directly. No translation layer required.None of this is new in theory. Vibe coding opened the gates of software creation to millions. That was aspiration. When I shared the data on how our engineers doubled throughput, shifted from coding to validation, brought design upfront for rapid experimentation, it was still an engineering story. What changed is that the theory became practice. Here’s how it actually played out.The bottleneck movedWhen we went AI-first in 2025, implementation cost collapsed. Agents took over scaffolding, tests, and the repetitive glue code that used to eat half the sprint. Cycle times dropped from weeks to days, from days to hours. Engineers started thinking less in files and functions and more in architecture, constraints, and execution plans.But once engineering capacity stopped being the bottleneck, we noticed something: Decision velocity was. All the coordination mechanisms we’d built to protect engineering time (specs, tickets, handoffs, backlog grooming) were now the slowest part of the system. We were optimizing for a constraint that no longer existed.What happens when building is cheaper than coordinationWe started asking a different question: What would it look like if the people closest to the intent could ship the software directly?PMs already think in specifications. Designers already define structure, layout, and behavior. They don’t think in syntax. They think in outcomes. When the cost of turning intent into working software dropped far enough, these roles didn’t need to “learn to code.” The cost of implementation simply fell to their level.I asked one of our PMs, Dmitry, to describe what changed from his perspective. He told me: “While agents are generating tasks in Zenflow, there’s a few minutes of idle time. Just dead air. I wanted to build a small game, something to interact with while you wait.”If you’ve ever run a product team, you know this kind of idea. It doesn’t move a KPI. It’s impossible to justify in a prioritization meeting. It gets deferred forever. But it adds personality. It makes the product feel like someone cared about the small details. These are exactly the things that get optimized out of every backlog grooming session, and exactly the things users remember.He built it in a day. In the past, that idea would have died in a prioritization spreadsheet. Not because it was bad, but because the cost of implementation made it irrational to pursue. When that cost drops to near zero, the calculus changes completely.Shipping became cheaper than explainingAs more people started building directly, entire layers of process quietly vanished. Fewer tickets. Fewer handoffs. Fewer “can you explain what you mean by…” conversations. Fewer lost-in-translation moments.For a meaningful class of tasks, it became faster to just build the thing than to describe what you wanted and wait for someone else to build it. Think about that for a second. Every modern software organization is structured around the assumption that implementation is the expensive part. When that assumption breaks, the org has to change with it.Our designer fixing the plugin UI is a perfect example. The old workflow (screenshot the problem, file a ticket, explain the gap between intent and implementation, wait for a sprint slot, review the result, request adjustments) existed entirely to protect engineering bandwidth. When the person with the design intuition can act on it directly, that whole stack disappears. Not because we eliminated process for its own sake, but because the process was solving a problem that no longer existed.The compounding effectHere’s what surprised me most: It compounds.When PMs build their own ideas, their specifications get sharper, because they now understand what the agent needs to execute well. Sharper specs produce better agent output. Better output means fewer iteration cycles. We’re seeing velocity compound week over week, not just because the models improved, but because the people using them got closer to the work.Dmitry put it well: The feedback loop between intent and outcome went from weeks to minutes. When you can see the result of your specification immediately, you learn what precision the system needs, and you start providing it instinctively.There’s a second-order effect that’s harder to measure but impossible to miss: Ownership. People stop waiting. They stop filing tickets for things they could just fix. “Builder” stopped being a job title. It became the default behavior.What this means for the industryA lot of the “everyone can code” narrative last year was theoretical, or focused on solo founders and tiny teams. What we experienced is different. We have ~50 engineers working in a complex brownfield codebase: Multiple surfaces and programming languages, enterprise integrations, the full weight of a real production system. I don’t think we’re unique. I think we’re early. And with each new generation of models, the gap between who can build and who can’t is closing faster than most organizations realize. Every software company is about to discover that their PMs and designers are sitting on unrealized building capacity, blocked not by skill, but by the cost of implementation. As that cost continues to fall, the organizational implications are profound.We started with an intent to accelerate software engineering. What we’re becoming is something different: A company where everyone ships.Andrew Filev is founder and CEO of Zencoder.

Many people have tried AI tools and walked away unimpressed. I get it — many demos promise magic, but in practice, the results can feel underwhelming.That’s why I want to write this not as a futurist prediction, but from lived experience. Over the past six months, I turned my engineering organization AI-first. I’ve shared before about the system behind that transformation — how we built the workflows, the metrics, and the guardrails. Today, I want to zoom out from the mechanics and talk about what I’ve learned from that experience — about where our profession is heading when software development itself turns inside out. Before I do, a couple of numbers to illustrate the scale of change. Subjectively, it feels that we are moving twice as fast. Objectively, here’s how the throughput evolved. Our total engineering team headcount floated from 36 at the beginning of the year to 30. So you get ~170% throughput on ~80% headcount, which matches the subjective ~2x. Zooming in, I picked a couple of our senior engineers who started the year in a more traditional software engineering process and ended it in the AI-first way. [The dips correspond to vacations and off-sites]:Note that our PRs are tied to JIRA tickets, and the average scope of those tickets didn’t change much through the year, so it’s as good a proxy as the data can give us. Qualitatively, looking at the business value, I actually see even higher uplift. One reason is that, as we started last year, our quality assurance (QA) team couldn’t keep up with our engineers’ velocity. As the company leader, I wasn’t happy with the quality of some of our early releases. As we progressed through the year, and tooled our AI workflows to include writing unit and end-to-end tests, our coverage improved, the number of bugs dropped, users became fans, and the business value of engineering work multiplied.From big design to rapid experimentationBefore AI, we spent weeks perfecting user flows before writing code. It made sense when change was expensive. Agile helped, but even then, testing multiple product ideas was too costly.Once we went AI-first, that trade-off disappeared. The cost of experimentation collapsed. An idea could go from whiteboard to a working prototype in a day: From idea to AI-generated product requirements document (PRD), to AI-generated tech spec, to AI-assisted implementation. It manifested itself in some amazing transformations. Our website—central to our acquisition and inbound demand—is now a product-scale system with hundreds of custom components, all designed, developed, and maintained directly in code by our creative director. Now, instead of validating with slides or static prototypes, we validate with working products. We test ideas live, learn faster, and release major updates every other month, a pace I couldn’t imagine three years ago.For example, Zen CLI was first written in Kotlin, but then we changed our mind and moved it to TypeScript with no release velocity lost.Instead of mocking the features, our UX designers and project managers vibe code them. And when the release-time crunch hit everyone, they jumped into action and fixed dozens of small details with production-ready PRs to help us ship a great product. This included an overnight UI layout change.From coding to validationThe next shift came where I least expected it: Validation.In a traditional org, most people write code and a smaller group tests it. But when AI generates much of the implementation, the leverage point moves. The real value lies in defining what “good” looks like — in making correctness explicit.We support 70-plus programming languages and countless integrations. Our QA engineers have evolved into system architects. They build AI agents that generate and maintain acceptance tests directly from requirements. And those agents are embedded into the codified AI workflows that allow us to achieve predictable engineering outcomes by using a system.This is what “shift left” really means. Validation isn’t a stand-alone function, it’s an integral part of the production process. If the agent can’t validate it’s work, it can’t be trusted to generate production code. For QA professionals, this is a moment of reinvention, where, with the right upskilling, their work becomes a critical enabler and accelerator of the AI adoption. Product managers, tech leads, and data engineers now share this responsibility as well, because defining correctness has become a cross-functional skill, not a role confined to QA.From diamond to double funnelFor decades, software development followed a “diamond” shape: A small product team handed off to a large engineering team, then narrowed again through QA.Today, that geometry is flipping. Humans engage more deeply at the beginning — defining intent, exploring options — and again at the end, validating outcomes. The middle, where AI executes, is faster and narrower.It’s not just a new workflow; it’s a structural inversion.The model looks less like an assembly line and more like a control tower. Humans set direction and constraints, AI handles execution at speed, and people step back in to validate outcomes before decisions land in production.Engineering at a higher level of abstractionEvery major leap in software raised our level of abstraction — from punch cards to high-level programming languages, from hardware to cloud. AI is the next step. Our engineers now work at a meta-layer: Orchestrating AI workflows, tuning agentic instructions and skills, and defining guardrails. The machines build; the humans decide what and why.Teams now routinely decide when AI output is safe to merge without review, how tightly to bound agent autonomy in production systems, and what signals actually indicate correctness at scale, decisions that simply didn’t exist before.And that’s the paradox of AI-first engineering — it feels less like coding, and more like thinking. Welcome to the new era of human intelligence, powered by AI.Andrew Filev is founder and CEO of Zencoder

Processing 200,000 tokens through a large language model is expensive and slow: the longer the context, the faster the costs spiral. Researchers at Tsinghua University and Z.ai have built a technique called IndexCache that cuts up to 75% of the redundant computation in sparse attention models, delivering up to 1.82x faster time-to-first-token and 1.48x faster generation throughput at that context length.The technique applies to models using the DeepSeek Sparse Attention architecture, including the latest DeepSeek and GLM families. It can help enterprises provide faster user experiences for production-scale, long-context models, a capability already proven in preliminary tests on the 744-billion-parameter GLM-5 model.The DSA bottleneckLarge language models rely on the self-attention mechanism, a process where the model computes the relationship between every token in its context and all the preceding ones to predict the next token.However, self-attention has a severe limitation. Its computational complexity scales quadratically with sequence length. For applications requiring extended context windows (e.g., large document processing, multi-step agentic workflows, or long chain-of-thought reasoning), this quadratic scaling leads to sluggish inference speeds and significant compute and memory costs.Sparse attention offers a principled solution to this scaling problem. Instead of calculating the relationship between every token and all preceding ones, sparse attention optimizes the process by having each query select and attend to only the most relevant subset of tokens.DeepSeek Sparse Attention (DSA) is a highly efficient implementation of this concept, first introduced in DeepSeek-V3.2. To determine which tokens matter most, DSA introduces a lightweight “lightning indexer module” at every layer of the model. This indexer scores all preceding tokens and selects a small batch for the main core attention mechanism to process. By doing this, DSA slashes the heavy core attention computation from quadratic to linear, dramatically speeding up the model while preserving output quality.But the researchers identified a lingering flaw: the DSA indexer itself still operates at a quadratic complexity at every single layer. Even though the indexer is computationally cheaper than the main attention process, as context lengths grow, the time the model spends running these indexers skyrockets. This severely slows down the model, especially during the initial “prefill” stage where the prompt is first processed.Caching attention with IndexCacheTo solve the indexer bottleneck, the research team discovered a crucial characteristic of how DSA models process data. The subset of important tokens an indexer selects remains remarkably stable as data moves through consecutive transformer layers. Empirical tests on DSA models revealed that adjacent layers share between 70% and 100% of their selected tokens.To capitalize on this cross-layer redundancy, the researchers developed IndexCache. The technique partitions the model’s layers into two categories. A small number of full (F) layers retain their indexers, actively scoring the tokens and choosing the most important ones to cache. The rest of the layers become shared (S), performing no indexing and reusing the cached indices from the nearest preceding F layer.During inference, the model simply checks the layer type. If it reaches an F layer, it calculates and caches fresh indices. If it is an S layer, it skips the math and copies the cached data.There is a wide range of optimization techniques that try to address the attention bottleneck by compressing the KV cache, where the computed attention values are stored. Instead of shrinking the memory footprint like standard KV cache compression, IndexCache attacks the compute bottleneck. “IndexCache is not a traditional KV cache compression or sharing technique,” Yushi Bai, co-author of the paper, told VentureBeat. “It eliminates this redundancy by reusing indices across layers, thereby reducing computation rather than just memory footprint. It is complementary to existing approaches and can be combined with them.”The researchers developed two deployment approaches for IndexCache. (It is worth noting that IndexCache only applies to models that use the DSA architecture, such as the latest DeepSeek models and the latest family of GLM models.)For developers working with off-the-shelf DSA models where retraining is unfeasible or too expensive, they created a training-free method relying on a “greedy layer selection” algorithm. By running a small calibration dataset through the model, this algorithm automatically determines the optimal placement of F and S layers without any weight updates. Empirical evidence shows that the greedy algorithm can safely remove 75% of the indexers while matching the downstream performance of the original model.For teams pre-training or heavily fine-tuning their own foundation models, the researchers propose a training-aware version that optimizes the network parameters to natively support cross-layer sharing. This approach introduces a “multi-layer distillation loss” during training. It forces each retained indexer to learn how to select a consensus subset of tokens that will be highly relevant for all the subsequent layers it serves.Real-world speedups on production modelsTo test the impact of IndexCache, the researchers applied it to the 30-billion-parameter GLM-4.7 Flash model and compared it against the standard baseline.At a 200K context length, removing 75% of the indexers slashed the prefill latency from 19.5 seconds down to just 10.7 seconds, delivering a 1.82x speedup. The researchers note these speedups are expected to be even greater in longer contexts.During the decoding phase, where the model generates its response, IndexCache boosted per-request throughput from 58 tokens per second to 86 tokens per second at the 200K context mark, yielding a 1.48x speedup. When the server’s memory is fully saturated with requests, total decode throughput jumped by up to 51%.For enterprise teams, these efficiency gains translate directly into cost savings. “In terms of ROI, IndexCache provides consistent benefits across scenarios, but the gains are most noticeable in long-context workloads such as RAG, document analysis, and agentic pipelines,” Bai said. “In these cases, we observe at least an approximate 20% reduction in deployment cost and similar improvements in user-perceived latency.” He added that for very short-context tasks, the benefits hover around 5%.Remarkably, these efficiency gains did not compromise reasoning capabilities. Using the training-free approach to eliminate 75% of indexers, the 30B model matched the original baseline’s average score on long-context benchmarks, scoring 49.9 against the original 50.2. On the highly complex AIME 2025 math reasoning benchmark, the optimized model actually outperformed the original baseline, scoring 92.6 compared to 91.0.The team also ran preliminary experiments on the production-scale 744-billion-parameter GLM-5 model. They found that eliminating 75% of its indexers with the training-free method yielded at least a 1.3x speedup on contexts over 100K tokens. At the same time, the model maintained a nearly identical quality average on long-context tasks.Getting IndexCache into productionFor development teams wanting to implement the training-free approach today, the process is straightforward but requires careful setup. While the greedy search algorithm automatically finds the optimal layer configuration, the quality of that configuration depends on the data it processes.“We recommend using domain-specific data as a calibration set so that the discovered layer-sharing pattern aligns with real workloads,” Bai said.Once calibrated, the optimization is highly accessible for production environments. Open-source patches are already available on GitHub for major serving engines. “Integration is relatively straightforward — developers can apply the patch to existing inference stacks, such as vLLM or SGLang, and enable IndexCache with minimal configuration changes,” Bai said.While IndexCache provides an immediate fix for today’s compute bottlenecks, its underlying philosophy points to a broader shift in how the AI industry will approach model design.“Future foundation models will likely be architected with downstream inference constraints in mind from the beginning,” Bai concluded. “This means designs that are not only scalable in terms of model size, but also optimized for real-world throughput and latency, rather than treating these as post-hoc concerns.”

Intercom is taking an unusual gamble for a legacy software company: building its own AI model.The 15-year-old, Dublin, Ireland-based massive customer service platform announced Fin Apex 1.0 on Thursday, a small, purpose-built AI model that the company claims outperforms leading frontier models from OpenAI and Anthropic on the metrics that matter most for customer support. The model powers Intercom’s existing Fin AI agent, which already handles over one million customer conversations weekly.According to benchmarks shared with VentureBeat, Fin Apex 1.0 achieves a 73.1% resolution rate—the percentage of customer issues fully resolved without human intervention—compared to 71.1% for both GPT-5.4 and Claude Opus 4.5, and 69.6% for Claude Sonnet 4.6. That roughly 2 percentage point margin may sound modest, but it’s wider than the typical gap between successive generations of frontier models.”If you’re running large service operations at scale and you’ve got 10 million customers or a billion dollars in revenue, a delta of 2% or 3% is a really large amount of customers and interactions and revenue,” Intercom CEO Eoghan McCabe told VentureBeat in a video call interview earlier this week.The model also shows significant improvements in speed and accuracy. Fin Apex delivers responses in 3.7 seconds—0.6 seconds faster than the next-fastest competitor—and demonstrates a 65% reduction in hallucinations compared to Claude Sonnet 4.6. Perhaps most striking for enterprise buyers: it runs at roughly one-fifth the cost of using frontier models directly, and is included in Intercom’s existing “per-outcome”-based pricing structure for its existing customer plans.What’s the base model? Does it even matter?But there’s a catch. When asked to specify which base model Apex was built on—and its parameter size—Intercom declined.”We’re not sharing the base model we used for Apex 1.0—for competitive reasons and also because we plan to switch base models over time,” a company spokesperson told VentureBeat. The company would only confirm that the model is “in the size of hundreds of millions of parameters.”That’s a notably small model. For comparison, Meta’s Llama 3.1 ranges from 8 billion to 405 billion parameters; even efficient open-weights models like Mistral 7B dwarf the sub-billion scale Intercom describes. Whether Apex’s performance claims hold up against that context—or whether the benchmarks reflect optimizations possible only in narrow, domain-specific applications—remains an open question.Intercom says it learned from the backlash AI coding startup Cursor faced when critics accused the coding assistant of burying the fact that its Composer 2 model was built on fine-tuned open-weights models rather than proprietary technology. But the lesson Intercom drew may not satisfy skeptics: the company is transparent that it used an open-weights base, just not which one.”We are very transparent that we have” used an open-weights model, the spokesperson said. Yet declining to name the model while claiming transparency is a contradiction that will likely draw scrutiny—particularly as more companies tout “proprietary” AI that amounts to post-trained open-source foundations.Post-training as the new frontierIntercom’s argument is that the base model simply doesn’t matter much anymore.”Pre-training is kind of a commodity now,” McCabe said. “The frontier, if you will, is actually in post-training. Post-training is the hard part. You need proprietary data. You need proprietary sources of truth.”The company post-trained its chosen foundation using years of proprietary customer service data accumulated through Fin, which now resolves 2 million customer queries per week. That process involved more than just feeding transcripts into a model. Intercom built reinforcement learning systems grounded in real resolution outcomes, teaching the model what successful customer service actually looks like—the appropriate tone, judgment calls, conversational structure, and critically, how to recognize when an issue is truly resolved versus when a customer is still frustrated.”The generic models are trained on generic data on the internet. The specific models are trained on hyper-specific domain data,” McCabe explained. “It stands to reason therefore that the intelligence of the generic models is generic, and the intelligence of the specific models is domain-specific and therefore operates in a far superior way for that use case.”If McCabe is right that the magic is entirely in post-training, the reluctance to name the base becomes harder to justify. If the foundation is truly interchangeable, what competitive advantage does secrecy protect?A $100 million bet paying offThe announcement comes as Intercom’s AI-first pivot appears to be working. Fin is approaching $100 million in annual recurring revenue and growing at 3.5x, making it the fastest-growing segment of the company’s $400 million ARR business. Fin is projected to represent half of Intercom’s total revenue early next year.That trajectory represents a remarkable turnaround. When Fin launched, its resolution rate was just 23%. Today it averages 67% across customers, with some large enterprise deployments seeing rates as high as 75%.To make this happen, Intercom grew its AI team from roughly 6 researchers to 60 over the past three years—a significant investment for a company that McCabe admits was “in a really bad place” before its AI pivot. The average growth rate for public software companies sits around 11%; Intercom expects to hit 37% growth this year.”We’re by far the first in the category to train our own model,” McCabe said. “There’s no one else that’s going to have this for a year or more.”The speciation and specialization of AIMcCabe’s thesis aligns with a broader trend that Andrej Karpathy, former AI leader at Tesla and OpenAI, recently described as the “speciation” of AI models—a proliferation of specialized systems optimized for narrow tasks rather than general intelligence.Customer service, McCabe argues, is uniquely suited for this approach. It’s one of only two or three enterprise AI use cases that have found genuine economic traction so far, alongside coding assistants and potentially legal AI. That’s attracted over a billion dollars in venture funding to competitors like Decagon and Sierra—and made the space, in McCabe’s words, “ruthlessly competitive.”The question is whether domain-specific models represent a durable advantage or a temporary arbitrage that frontier labs will eventually close. McCabe believes the labs face structural limitations.”Maybe the future is that Anthropic has a big offering of many different specialized models. Maybe that’s what it looks like,” he said. “But the reality is that I don’t think the generic models are going to be able to keep up with the domain-specific models right now.”Beyond efficiency to experienceEarly enterprise AI adoption focused heavily on cost reduction—replacing expensive human agents with cheaper automated ones. But McCabe sees the conversation shifting toward experience quality.”Originally it was like, ‘Holy shit, we can actually do this for so much cheaper.’ And now they’re thinking, ‘Wait, no, we can give customers a far better experience,'” he said.The vision extends beyond simple query resolution. McCabe imagines AI agents that function as consultants—a shoe retailer’s bot that doesn’t just answer shipping questions but offers styling advice and shows customers how different options might look on them.”Customer service has always been pretty shit,” McCabe said bluntly. “Even the very best brands, you’re left waiting on a call, you’re bounced around different departments. There’s an opportunity now to provide truly perfect customer experience.”Pricing and availabilityFor existing Fin customers, the upgrade to Apex comes at no additional cost. Intercom confirmed that customer pricing remains unchanged—users continue to pay per outcome as before, at $0.99 per resolved interaction, and automatically benefit from the new model.Apex is not available as a standalone model or through an external API. It is accessible only through Fin, meaning businesses cannot license the model independently or integrate it into their own products. That constraint may limit Intercom’s ability to monetize the model beyond its existing customer base—but it also keeps the technology proprietary in a practical sense, regardless of what the underlying base model turns out to be.What’s nextIntercom plans to expand Fin beyond customer service into sales and marketing—positioning it as a direct competitor to Salesforce’s Agentforce vision, which aims to provide AI agents across the customer lifecycle.For the broader SaaS industry, Intercom’s move raises uncomfortable questions. If a 15-year-old customer service company can build a model that outperforms OpenAI and Anthropic in its domain, what does that mean for vendors still relying on generic API calls? And if “post-training is the new frontier,” as McCabe insists, will companies claiming breakthroughs face pressure to show their work—or continue hiding behind competitive secrecy while touting transparency?McCabe’s answer to the first question, laid out in a recent LinkedIn post, is stark: “If you can’t become an agent company, your CRUD app business has a diminishing future.”The answer to the second remains to be seen.

The enterprise voice AI market is in the middle of a land grab. ElevenLabs and IBM announced a collaboration just this week to bring premium voice capabilities into IBM’s watsonx Orchestrate platform. Google Cloud has been expanding its Chirp 3 HD voices. OpenAI continues to iterate on its own speech synthesis. And the market underpinning all of this activity is enormous — voice AI crossed $22 billion globally in 2026, with the voice AI agents segment alone projected to reach $47.5 billion by 2034, according to industry estimates.On Thursday morning, Mistral AI entered that fight with a fundamentally different proposition. The Paris-based AI startup released Voxtral TTS, what it calls the first frontier-quality, open-weight text-to-speech model designed specifically for enterprise use. Where every major competitor in the space operates a proprietary, API-first business — enterprises rent the voice, they don’t own it — Mistral is releasing the full model weights, inviting companies to download Voxtral TTS, run it on their own servers or even on a smartphone, and never send a single audio frame to a third party.It is a bet that the future of enterprise voice AI will not be shaped by whoever builds the best-sounding model, but by whoever gives companies the most control over it. And it arrives at a moment when Mistral, valued at $13.8 billion after a $2 billion Series C round led by Dutch chipmaker ASML last September, has been aggressively assembling the building blocks of a complete, enterprise-owned AI stack — from its Forge customization platform announced at Nvidia GTC earlier this month, to its AI Studio production infrastructure, to the Voxtral Transcribe speech-to-text model released just weeks ago.Voxtral TTS is the output layer that completes that picture, giving enterprises a speech-to-speech pipeline they can run end-to-end without relying on any external provider.”We see audio as a big bet and as a critical and maybe the only future interface with all the AI models,” Pierre Stock, Mistral’s vice president of science and the first employee hired at the company, said in an exclusive interview with VentureBeat. “This is something customers have been asking for.”A 3-billion-parameter model that fits on a laptop and runs six times faster than real-time speechThe technical specifications of Voxtral TTS read like a deliberate inversion of industry norms. Where most frontier TTS models are large and resource-intensive, Mistral built its model to be roughly three times smaller than what it calls the industry standard for comparable quality.The architecture comprises three components: a 3.4-billion-parameter transformer decoder backbone, a 390-million-parameter flow-matching acoustic transformer, and a 300-million-parameter neural audio codec that Mistral developed in-house. The system is built on top of Ministral 3B, the same pretrained backbone that powers the company’s Voxtral Transcribe model — a design choice that Stock described as emblematic of Mistral’s culture of efficiency and artifact reuse.In practice, the model achieves a time-to-first-audio of 90 milliseconds for a typical input and generates speech at approximately six times real-time speed. When quantized for inference, it requires roughly three gigabytes of RAM. Stock confirmed it can run on any laptop or smartphone, and even on older hardware it still operates in real time.”It’s a 3B model, so it can basically run on any laptop or any smartphone,” Stock told VentureBeat. “If you quantize it to infer, it’s actually three gigabytes of RAM. And you can run it on super old chips — it’s still going to be real time.”The model supports nine languages — English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, and Arabic — and can adapt to a custom voice with as little as five seconds of reference audio. Perhaps more remarkably, it demonstrates zero-shot cross-lingual voice adaptation without explicit training for that task.Stock illustrated this with a personal example: he can feed the model 10 seconds of his own French-accented voice, type a prompt in German, and the model will generate German speech that sounds like him — complete with his natural accent and vocal characteristics. For enterprises operating across borders, this capability unlocks cascaded speech-to-speech translation that preserves speaker identity, a feature that has obvious applications in customer support, sales, and internal communications for multinational organizations.Human evaluators preferred Voxtral over ElevenLabs nearly 70 percent of the time on voice customizationMistral is not being coy about which competitor it intends to displace. In human evaluations conducted by the company, Voxtral TTS achieved a 62.8 percent listener preference rate against ElevenLabs Flash v2.5 on flagship voices and a 69.9 percent preference rate in voice customization tasks. Mistral also claims the model performs at parity with ElevenLabs v3 — the company’s premium, higher-latency tier — on emotional expressiveness, while maintaining similar latency to the much faster Flash model.The evaluation methodology involved a comparative side-by-side test across all nine supported languages. Using two recognizable voices in their native dialects for each language, three annotators performed preference tests on naturalness, accent adherence, and acoustic similarity to the original reference. Mistral says Voxtral TTS widened the quality gap to ElevenLabs v2.5 Flash especially in zero-shot multilingual custom voice settings, highlighting what the company calls the “instant customizability” of the model.ElevenLabs remains widely regarded as the benchmark for raw voice quality. Its Eleven v3 model has been described by multiple independent reviewers as the gold standard for emotionally nuanced AI speech. But ElevenLabs operates as a closed platform with tiered subscription pricing that scales from around $5 per month at the starter level to over $1,300 per month for business plans. It does not release model weights.Mistral’s pitch is that enterprises shouldn’t have to choose between quality and control — and that at scale, the economics of an open-weight model are dramatically more favorable.”What we want to underline is that we’re faster and cheaper as well — and open source,” Stock told VentureBeat. “When something is open source and cheap, people adopt it and people build on it.”He framed the cost argument in terms that resonate with CTOs managing AI budgets: “AI is a transformative technology, but it has a cost. When you want to scale and have impact on a large business, that cost matters. And what we allow is to scale seamlessly while minimizing the cost and maximizing the accuracy.”Why Mistral thinks enterprises will want to own their voice AI rather than rent itTo understand why Mistral is entering text-to-speech now, you have to understand the broader strategic architecture the company has been building for the past year. While OpenAI and Anthropic have captured the imagination of consumers, Mistral has quietly assembled what may be the most comprehensive enterprise AI platform in Europe — and increasingly, globally.CEO Arthur Mensch has said the company is on track to surpass $1 billion in annual recurring revenue this year, according to TechCrunch’s reporting on the Forge launch. The Financial Times has reported that Mistral’s annualized revenue run rate surged from $20 million to over $400 million within a single year. That growth has been powered by more than 100 major enterprise customers and a consistent thesis: companies should own their AI infrastructure, not rent it.Voxtral TTS is the latest expression of that thesis, applied to what may be the most sensitive category of enterprise data there is. Voice recordings capture not just words but emotion, identity, and intent. They carry legal, regulatory, and reputational weight that text data often does not. For industries like financial services, healthcare, and government — all key Mistral verticals — sending voice data to a third-party API introduces risks that many compliance teams are unwilling to accept.Stock made the data sovereignty argument forcefully. “Since the models are open weights, we have no trouble and no problem actually giving the weights to the enterprise and helping them customize the models,” he said. “We don’t see the weights anymore. We don’t see the data. We see nothing. And you are fully controlled.”That message has particular resonance in Europe, where concern about technological dependence on American cloud providers has intensified throughout 2026. The EU currently sources more than 80 percent of its digital services from foreign providers, most of them American. Mistral has positioned itself as the answer to that anxiety — the only European frontier AI developer with the scale and technical capability to offer a credible alternative.Voice agents are the enterprise use case that makes Mistral’s full AI stack click into placeVoxtral TTS is the final piece in a pipeline Mistral has been methodically assembling. Voxtral Transcribe handles speech-to-text. Mistral’s language models — from Mistral Small to Mistral Large — provide the reasoning layer. Forge allows enterprises to customize any of these models on their own data. AI Studio provides the production infrastructure for observability, governance, and deployment. And Mistral Compute offers the underlying GPU resources.Together, these pieces form what Stock described as a “full AI stack, fully controllable and customizable” for the enterprise. Voice agents — AI systems that can listen to a customer, understand what they need, reason about the answer, and respond in natural-sounding speech — are the use case that ties all of these layers together.The applications Mistral envisions span customer support, where voice agents can route and resolve queries with brand-appropriate speech; sales and marketing, where a single voice can work across markets through cross-lingual emulation; real-time translation for cross-border operations; and even interactive storytelling and game design, where emotion-steering can control tone and personality.Stock was most animated when discussing how Voxtral TTS fits into the broader agentic AI trend that has dominated enterprise technology discussions in 2026. “We are totally building for a world in which audio is a natural interface, in particular for agents to which you can delegate work — extensions of yourself,” he said. He described a scenario in which a user starts planning a vacation on a computer, commutes to work, and then picks up the workflow on a phone simply by asking for an update by voice.”To make that happen, you need a model you can trust, you need a model that’s super efficient and super cheap to run — otherwise you won’t use it for long — and you need a model that sounds super conversational and that you can interrupt at any time,” Stock said.That emphasis on interruptibility and real-time responsiveness reflects a broader insight about voice interfaces that distinguishes them from text. A chatbot can take two or three seconds to respond without breaking the user experience. A voice agent cannot. The 90-millisecond time-to-first-audio that Voxtral TTS achieves is not just a benchmark number — it is the threshold between a voice interaction that feels natural and one that feels robotic.Mistral’s open-weight approach aligns with a broader industry shift that even Nvidia is backingMistral’s decision to release Voxtral TTS with open weights is consistent with a movement that has been gathering momentum across the AI industry. At Nvidia GTC earlier this month, Nvidia CEO Jensen Huang declared that “proprietary versus open is not a thing — it’s proprietary and open.” Nvidia announced the Nemotron Coalition, a first-of-its-kind collaboration of model builders working to advance open frontier-level foundation models, with Mistral as a founding member. The first project from that coalition will be a base model codeveloped by Mistral AI and Nvidia.For Mistral, open weights serve a dual commercial purpose. They drive adoption — developers and enterprises can experiment without friction or commitment — while the company monetizes through its platform services, customization offerings, and managed infrastructure. The model is available to test in Mistral Studio and through the company’s API, but the strategic play is to become embedded in enterprise voice pipelines as an owned asset, not a metered service.This mirrors the playbook that worked for Mistral’s language models. As Mensch told CNBC in February, “AI is making us able to develop software at the speed of light,” predicting that “more than half of what’s currently being bought by IT in terms of SaaS is going to shift to AI.” He described a “replatforming” taking place across enterprise technology, with businesses looking to replace legacy software systems with AI-native alternatives. An open-weight voice model that enterprises can customize and deploy on their own terms fits naturally into that narrative.Mistral signals that end-to-end audio AI is where the company is headed nextWhen asked what comes after Voxtral TTS, Stock outlined two directions. The first is expanding language and dialect support, with particular attention to cultural nuance. “It’s not the same to speak French in Paris than to speak French in Canada, in Montreal,” he said. “We want to respect both cultures, and we want our models to perform in both contexts with all the cultural specifics.”The second direction is more ambitious: a fully end-to-end audio model that doesn’t just generate speech from text but understands the complete spectrum of human vocal communication.”We convey some meaning with the words we speak,” Stock said. “We actually convey way more with the intonation, the rhythm, and how we say it. When people talk about end-to-end audio, that’s what they mean — the model is able to pick up that you’re in a hurry, for instance, and will go for the fastest answer. The model will know that you’re joyful today and crack a joke. It’s super adaptive to you, and that’s where we want to go.”That vision — an AI that speaks naturally, listens with nuance, responds with emotional intelligence, and runs on a model small enough to fit in your pocket — is the frontier every major AI lab is racing toward. For now, Voxtral TTS gives Mistral a foundation to build on and enterprises a question they haven’t had to answer before: if you could own your voice AI stack outright, at lower cost and with competitive quality, why would you keep renting someone else’s?

As Large Language Models (LLMs) expand their context windows to process massive documents and intricate conversations, they encounter a brutal hardware reality known as the “Key-Value (KV) cache bottleneck.”Every word a model processes must be stored as a high-dimensional vector in high-speed memory. For long-form tasks, this “digital cheat sheet” swells rapidly, devouring the graphics processing unit (GPU) video random access memory (VRAM) system used during inference, and slowing the model performance down rapidly over time. But have no fear, Google Research is here: yesterday, the unit within the search giant released its TurboQuant algorithm suite — a software-only breakthrough that provides the mathematical blueprint for extreme KV cache compression, enabling a 6x reduction on average in the amount of KV memory a given model uses, and 8x performance increase in computing attention logits, which could reduce costs for enterprises that implement it on their models by more than 50%. The theoretically grounded algorithms and associated research papers are available now publicly for free, including for enterprise usage, offering a training-free solution to reduce model size without sacrificing intelligence.The arrival of TurboQuant is the culmination of a multi-year research arc that began in 2024. While the underlying mathematical frameworks—including PolarQuant and Quantized Johnson-Lindenstrauss (QJL)—were documented in early 2025, their formal unveiling today marks a transition from academic theory to large-scale production reality. The timing is strategic, coinciding with the upcoming presentations of these findings at the upcoming conferences International Conference on Learning Representations (ICLR 2026) in Rio de Janeiro, Brazil, and Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026) in Tangier, Morocco. By releasing these methodologies under an open research framework, Google is providing the essential “plumbing” for the burgeoning “Agentic AI” era: the need for massive, efficient, and searchable vectorized memory that can finally run on the hardware users already own. Already, it is believed to have an effect on the stock market, lowering the price of memory providers as traders look to the release as a sign that less memory will be needed (perhaps incorrect, given Jevons’ Paradox).The Architecture of Memory: Solving the Efficiency TaxTo understand why TurboQuant matters, one must first understand the “memory tax” of modern AI. Traditional vector quantization has historically been a “leaky” process. When high-precision decimals are compressed into simple integers, the resulting “quantization error” accumulates, eventually causing models to hallucinate or lose semantic coherence. Furthermore, most existing methods require “quantization constants”—meta-data stored alongside the compressed bits to tell the model how to decompress them. In many cases, these constants add so much overhead—sometimes 1 to 2 bits per number—that they negate the gains of compression entirely.TurboQuant resolves this paradox through a two-stage mathematical shield. The first stage utilizes PolarQuant, which reimagines how we map high-dimensional space. Rather than using standard Cartesian coordinates (X, Y, Z), PolarQuant converts vectors into polar coordinates consisting of a radius and a set of angles. The breakthrough lies in the geometry: after a random rotation, the distribution of these angles becomes highly predictable and concentrated. Because the “shape” of the data is now known, the system no longer needs to store expensive normalization constants for every data block. It simply maps the data onto a fixed, circular grid, eliminating the overhead that traditional methods must carry.The second stage acts as a mathematical error-checker. Even with the efficiency of PolarQuant, a residual amount of error remains. TurboQuant applies a 1-bit Quantized Johnson-Lindenstrauss (QJL) transform to this leftover data. By reducing each error number to a simple sign bit (+1 or -1), QJL serves as a zero-bias estimator. This ensures that when the model calculates an “attention score”—the vital process of deciding which words in a prompt are most relevant—the compressed version remains statistically identical to the high-precision original.Performance benchmarks and real-world reliabilityThe true test of any compression algorithm is the “Needle-in-a-Haystack” benchmark, which evaluates whether an AI can find a single specific sentence hidden within 100,000 words. In testing across open-source models like Llama-3.1-8B and Mistral-7B, TurboQuant achieved perfect recall scores, mirroring the performance of uncompressed models while reducing the KV cache memory footprint by a factor of at least 6x. This “quality neutrality” is rare in the world of extreme quantization, where 3-bit systems usually suffer from significant logic degradation.Beyond chatbots, TurboQuant is transformative for high-dimensional search. Modern search engines increasingly rely on “semantic search,” comparing the meanings of billions of vectors rather than just matching keywords. TurboQuant consistently achieves superior recall ratios compared to existing state-of-the-art methods like RabbiQ and Product Quantization (PQ), all while requiring virtually zero indexing time. This makes it an ideal candidate for real-time applications where data is constantly being added to a database and must be searchable immediately. Furthermore, on hardware like NVIDIA H100 accelerators, TurboQuant’s 4-bit implementation achieved an 8x performance boost in computing attention logs, a critical speedup for real-world deployments.Rapt community reactionThe reaction on X, obtained via a Grok search, included a mixture of technical awe and immediate practical experimentation. The original announcement from @GoogleResearch generated massive engagement, with over 7.7 million views, signaling that the industry was hungry for a solution to the memory crisis.Within 24 hours of the release, community members began porting the algorithm to popular local AI libraries like MLX for Apple Silicon and llama.cpp.Technical analyst @Prince_Canuma shared one of the most compelling early benchmarks, implementing TurboQuant in MLX to test the Qwen3.5-35B model. Across context lengths ranging from 8.5K to 64K tokens, he reported a 100% exact match at every quantization level, noting that 2.5-bit TurboQuant reduced the KV cache by nearly 5x with zero accuracy loss. This real-world validation echoed Google’s internal research, proving that the algorithm’s benefits translate seamlessly to third-party models.Other users focused on the democratization of high-performance AI. @NoahEpstein_ provided a plain-English breakdown, arguing that TurboQuant significantly narrows the gap between free local AI and expensive cloud subscriptions. He noted that models running locally on consumer hardware like a Mac Mini “just got dramatically better,” enabling 100,000-token conversations without the typical quality degradation. Similarly, @PrajwalTomar_ highlighted the security and speed benefits of running “insane AI models locally for free,” expressing “huge respect” for Google’s decision to share the research rather than keeping it proprietary.Market impact and the future of hardwareThe release of TurboQuant has already begun to ripple through the broader tech economy. Following the announcement on Tuesday, analysts observed a downward trend in the stock prices of major memory suppliers, including Micron and Western Digital. The market’s reaction reflects a realization that if AI giants can compress their memory requirements by a factor of six through software alone, the insatiable demand for High Bandwidth Memory (HBM) may be tempered by algorithmic efficiency.As we move deeper into 2026, the arrival of TurboQuant suggests that the next era of AI progress will be defined as much by mathematical elegance as by brute force. By redefining efficiency through extreme compression, Google is enabling “smarter memory movement” for multi-step agents and dense retrieval pipelines. The industry is shifting from a focus on “bigger models” to “better memory,” a change that could lower AI serving costs globally.Strategic considerations for enterprise decision-makersFor enterprises currently using or fine-tuning their own AI models, the release of TurboQuant offers a rare opportunity for immediate operational improvement. Unlike many AI breakthroughs that require costly retraining or specialized datasets, TurboQuant is training-free and data-oblivious. This means organizations can apply these quantization techniques to their existing fine-tuned models—whether they are based on Llama, Mistral, or Google’s own Gemma—to realize immediate memory savings and speedups without risking the specialized performance they have worked to build.From a practical standpoint, enterprise IT and DevOps teams should consider the following steps to integrate this research into their operations:Optimize Inference Pipelines: Integrating TurboQuant into production inference servers can reduce the number of GPUs required to serve long-context applications, potentially slashing cloud compute costs by 50% or more.Expand Context Capabilities: Enterprises working with massive internal documentation can now offer much longer context windows for retrieval-augmented generation (RAG) tasks without the massive VRAM overhead that previously made such features cost-prohibitive.Enhance Local Deployments: For organizations with strict data privacy requirements, TurboQuant makes it feasible to run highly capable, large-scale models on on-premise hardware or edge devices that were previously insufficient for 32-bit or even 8-bit model weights.Re-evaluate Hardware Procurement: Before investing in massive HBM-heavy GPU clusters, operations leaders should assess how much of their bottleneck can be resolved through these software-driven efficiency gains.Ultimately, TurboQuant proves that the limit of AI isn’t just how many transistors we can cram onto a chip, but how elegantly we can translate the infinite complexity of information into the finite space of a digital bit. For the enterprise, this is more than just a research paper; it is a tactical unlock that turns existing hardware into a significantly more powerful asset.

Enterprise data teams moving agentic AI into production are hitting a consistent failure point at the data tier. Agents built across a vector store, a relational database, a graph store and a lakehouse require sync pipelines to keep context current. Under production load, that context goes stale. Oracle, whose database infrastructure runs the transaction systems of 97% of Fortune Global 100 companies by the company’s own count, is now making a direct architectural argument that the database is the right place to fix that problem.Oracle this week announced a set of agentic AI capabilities for Oracle AI Database, built around a direct architectural counter-argument to that pattern. The core of the release is the Unified Memory Core, a single ACID (Atomicity, Consistency, Isolation, and Durability)-transactional engine that processes vector, JSON, graph, relational, spatial and columnar data without a sync layer. Alongside that, Oracle announced Vectors on Ice for native vector indexing on Apache Iceberg tables, a standalone Autonomous AI Vector Database service and an Autonomous AI Database MCP Server for direct agent access without custom integration code.The news isn’t just that Oracle is adding new features, it’s about the world’s largest database vendor realizing that things have changed in the AI world that go beyond what its namesake database was providing.”As much as I’d love to tell you that everybody stores all their data in an Oracle database today — you and I live in the real world,” Maria Colgan, Vice President, Product Management for Mission-Critical Data and AI Engines, at Oracle told VentureBeat. “We know that that’s not true.”Four capabilities, one architectural bet against the fragmented agent stackOracle’s release spans four interconnected capabilities. Together they form the architectural argument that a converged database engine is a better foundation for production agentic AI than a stack of specialized tools.Unified Memory Core. Agents reasoning across multiple data formats simultaneously — vector, JSON, graph, relational, spatial — require sync pipelines when those formats live in separate systems. The Unified Memory Core puts all of them in a single ACID-transactional engine. Under the hood it is an API layer over the Oracle database engine, meaning ACID consistency applies across every data type without a separate consistency mechanism.

“By having the memory live in the same place that the data does, we can control what it has access to the same way we would control the data inside the database,” Colgan explained.Vectors on Ice. For teams running data lakehouse architectures on the open-source Apache Iceberg table format, Oracle now creates a vector index inside the database that references the Iceberg table directly. The index updates automatically as the underlying data changes and works with Iceberg tables that are managed by Databricks and Snowflake. Teams can combine Iceberg vector search with relational, JSON, spatial or graph data stored inside Oracle in a single query.Autonomous AI Vector Database. A fully managed, free-to-start vector database service built on the Oracle 26ai engine. The service is designed as a developer entry point with a one-click upgrade path to full Autonomous AI Database when workload requirements grow.Autonomous AI Database MCP Server. Lets external agents and MCP clients connect to Autonomous AI Database without custom integration code. Oracle’s row-level and column-level access controls apply automatically when an agent connects, regardless of what the agent requests.

“Even though you are making the same standard API call you would make with other platforms, the privileges that user has continued to kick in when the LLM is asking those questions,” Colgan said.Standalone vector databases are a starting point, not a destinationOracle’s Autonomous AI Vector Database enters a market occupied by purpose-built vector services including Pinecone, Qdrant and Weaviate. The distinction Oracle is drawing is about what happens when vector alone is not enough.”Once you are done with vectors, you do not really have an option,” Steve Zivanic, Global Vice President, Database and Autonomous Services, Product Marketing at Oracle, told VentureBeat. “With this, you can get graph, spatial, time series — whatever you may need. It is not a dead end.”Holger Mueller, principal analyst at Constellation Research, said that the architectural argument is credible precisely because other vendors cannot make it without moving data first. Other database vendors require transactional data to move to a data lake before agents can reason across it. Oracle’s converged legacy, in his view, gives it a structural advantage that is difficult to replicate without a ground-up rebuild.Not everyone sees the feature set as differentiated. Steven Dickens, CEO and principal analyst at HyperFRAME Research, told VentureBeat that vector search, RAG integration and Apache Iceberg support are now standard requirements across enterprise databases — Postgres, Snowflake and Databricks all offer comparable capabilities. “Oracle’s move to label the database itself as an AI Database is primarily a rebranding of its converged database strategy to match the current hype cycle,” Dickens said. In his view the real differentiation Oracle is claiming is not at the feature level but at the architectural level — and the Unified Memory Core is where that argument either holds or falls apart.Where enterprise agent deployments actually break downThe four capabilities Oracle shipped this week are a response to a specific and well-documented production failure mode. Enterprise agent deployments are not breaking down at the model layer. They are breaking down at the data layer, where agents built across fragmented systems hit sync latency, stale context and inconsistent access controls the moment workloads scale.Matt Kimball, vice president and principal analyst at Moor Insights and Strategy, told VentureBeat the data layer is where production constraints surface first. “The struggle is running them in production,” Kimball said. “The gap is seen almost immediately at the data layer — access, governance, latency and consistency. These all become constraints.”Dickens frames the core mismatch as a stateless-versus-stateful problem. Most enterprise agent frameworks store memory as a flat list of past interactions, which means agents are effectively stateless while the databases they query are stateful. The lag between the two is where decisions go wrong.

“Data teams are exhausted by fragmentation fatigue,” Dickens said. “Managing a separate vector store, graph database and relational system just to power one agent is a DevOps nightmare.”That fragmentation is precisely what Oracle’s Unified Memory Core is designed to eliminate. The control plane question follows directly.

“In a traditional application model, control lives in the app layer,” Kimball said. “With agentic systems, access control breaks down pretty quickly because agents generate actions dynamically and need consistent enforcement of policy. By pushing all that control into the database, it can all be applied in a more uniform way.”What this means for enterprise data teamsThe question of where control lives in an enterprise agentic AI stack is not settled.

Most organizations are still building across fragmented systems, and the architectural decisions being made now — which engine anchors agent memory, where access controls are enforced, how lakehouse data gets pulled into agent context — will be difficult to undo at scale.The distributed data challenge is still the real test.

“Data is increasingly distributed across SaaS platforms, lakehouses and event-driven systems, each with its own control plane and governance model,” Kimball said. “The opportunity now is extending that model across the broader, more distributed data estates that define most enterprise environments today.”