AI21 Labs

Excited to see our work featured in Jensen Huang’s keynote. As he put it: “Now you can turn your proprietary data into sovereign intelligence.” That’s exactly the vision we’re building at AI21 Labs. We believe every company will become an AI company - and that means owning and customizing its own AI, not just consuming it. Proud of the team’s tremendous work, and excited to partner with NVIDIA and industry leaders like Gong, Deel, and Alice (Formerly ActiveFence) to turn this into reality.

A rented frontier model makes you one of many. But an open model post-trained on your data, built for your workflows, and compounding with every interaction you feed back into it? That’s intelligence your competitors can never get. Read how 3 companies are building their AI edge with post-trained models from AI21. https://lnkd.in/e3zjm6PT

Over the past three weeks, our Forward-Deployed Researchers [FDRs] collaborated with NVIDIA, Gong, Deel, and Alice (Formerly ActiveFence) to post-train NVIDIA’s recently released Nemotron 3.5 Super VL open model to deliver specialized intelligence across three different use cases. We believe that the companies that get the most out of AI won't all be using the same intelligence. More and more companies are moving towards AI independence, aiming to own their models for the right reasons: lower cost, control over their data and models, and a proprietary edge a rented model can't offer. Few get there alone. Post-training that holds up in production still takes expertise that sits in a handful of labs. We were glad to partner on this path with Gong, Deel and Alice, training specialized models that match or surpass frontier quality – at a fraction of the cost – across data extraction, policy QA and attack resilience use cases. Full results and methods: https://lnkd.in/e3zjm6PT

Adding a verifier elevates every agentic AI system - but frontier verifiers are expensive. The good news: you can train a dedicated verifier and serve it at a fraction of the cost. In agentic search, the right answer is usually already in the pool. The system just doesn’t pick it. Majority voting can’t fix that. It simply ratifies whatever the crowd already believes, even when the crowd is wrong. What works: an independent verifier that re-researches every candidate and vetoes the bad ones. Only the survivors get voted on. On Google DeepMind’s FACTS-Search: • Frontier verifier: 83.3 → 93.4 (past published SOTA of 89.4) • Our trained 8B verifier: 92.9 at 3.2× lower cost It also generalizes to a new benchmark and search tool it never saw during training. This is the kind of result I like most - not just throwing a bigger model at the problem, but finding a better way to decide which answer to trust. Great work by Guy Freund, Eran Goldstein, Niv Granot & Yuval Belfer. Full post + training details → https://lnkd.in/gqgUJsfx

Your cheap AI agent probably already found the right answer. It just didn’t pick it. That’s the thing we kept running into over the past few months, working on making AI agents better at web research. When an agent gets a question wrong, the instinct is to buy a smarter model. That works. It’s also the most expensive option on the table. But give it a few attempts and a correct answer usually shows up somewhere in the pool. It just doesn't get selected. Majority voting doesn't fix this either, because it can't overturn a consensus on a wrong answer, and cheap pools are where that happens most. So the bottleneck is selection, not generation. The real job is curating the candidate pool: dropping the incorrect candidates without discarding the correct ones. And that's a much easier problem than generating the answer in the first place. So instead of improving the answering, we added a second agent whose only job is to check: given a question and a proposed answer, go research it on the web, come back with a verdict — valid or not. It worked. State of the art on FACTS-Search, DeepMind’s web research benchmark. It was also brutally expensive: the judge re-researches every candidate, so it ended up costing more than generating the answers in the first place. So we trained our own small 8B judge. Two approaches failed: → Imitation alone — it learns to search, but the verdict comes out fuzzy. The lesson gets spread across thousands of tokens, while the actual decision lives in one. → Reinforcement learning alone — no competent starting policy to explore from, so it gets stuck and never converges. Imitation first, then RL. That worked. We kept nearly all the quality at 3.2x lower cost. How you run your models matters more than which model you run. Guy Freund Niv Granot Yuval Belfer AI21 Labs. Full write-up 👇 https://lnkd.in/dSdKeqw2

AI usage is only moving one way ⬆️ But your bill doesn't have to follow. AI21 Gateway is a drop-in endpoint that learns each agent's cost profile from its own traffic and cuts spend (up to 68% on coding agents) without harming what your agents deliver. One base-URL change. No SDK, no rewrite. Free during beta. Get access now.