1/ Today we're introducing Cua-S1-4B-0.2, the first multimodal decision model trained with RLOO on live computer-use tasks, using task-completion rewards. Text and multimodal adapters are available under Apache-2.0:

Computer-use agent experience = capability × robustness × efficiency. As frontier models climb higher on CUA benchmarks, efficiency becomes increasingly attrative, as highlighted by the recent release of #Jev#. With many intermediate steps in everyday computer-use tasks being “local,” fast System-1 models like #CUA-S1# and #Jev# can make agent execution much more efficient. Cool exploration by @francedot and @ddupont808 at @trycua!

Genius is 99% perspiration, and 1% inspiration. This is also true for how we trained Periodic Neon to beat frontier models, like Astra and Fable 5.1, on a scientific capability our labs care about. We trained Neon with reinforcement learning on our lab data using only 1.3K GPUs, orders of magnitude less than Astra. As we scale our labs and compute, we expect our models to tackle much more complex scientific problems.

Five months ago, I joined NeoCognition as an intern. I started by digging through docs just to figure out how to use our clusters. Over time, I learned by doing and getting feedback from my manager, and gradually took on more of the work. And today, I’m proud to share that work: ApprenticeBench, the first continual learning benchmark grounded in a realistic work environment. At its heart, ApprenticeBench asks: Can an AI agent do the same just like I did? Use a computer to find its way through messy offline docs, learn from online experience and feedback over long horizons, and grow into the job The answer is yes and no. Only Fable 5.1 and Astra, released last week, outperform our best human tester, marking a real step change! But they still fall behind in efficiency. Humans consolidate what they learn and get more efficient with practice, while agents still struggle to do the same. Though there’s still room for improvement, AI agents have crossed an important threshold of job readiness. I couldn’t be more excited to help build the infra and harness behind ApprenticeBench and get the chance to witness this milestone firsthand. Side fun observations as a safety/security researcher: Echo to recent OAI hugging face incidents, across agent traces, we observed attempts to escape the sandbox, reach the internet, hunt for the grader and other reward-seeking behaviors. Thus, we put substantial effort into hardening the infrastructure, properly scoping the boundaries to secure the execution and get trustworthy results.

Introducing Speculative Programmatic Tool Calling (sPTC)! A general class of technique for speculating on tool calls during code generation in a harness and queuing them early to overlap with token generation + REPL execution time. Blog:

Introducing GEN-1.5, a one-shot learner. It can learn new tasks in a few seconds. Show it what to do, and it generalizes. This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.

🍰 CAKE paper's out, the design bet: the compiler isn't a fixed black box the agent calls — it's part of the harness, and it's under evolution too. CAKE didn't inherit existing abstraction layer. no tile/layout abstractions: the vocabulary was distilled by agents from a corpus of production kernels. every pattern the agent couldn't express pushed new primitives into the IR, and the analyses to keep them checkable. every barrier/layout bug that kept coming back became a verifier rule. none of this can be designed up front. the IR has to co-evolve with the kernels, and the workload tells you what's missing, the corpus tells you if the fix broke anything. the best language for an agent is the one that tells you what's illegal, what's slow, and which decision might made it faster.

I’ve always been bothered by how much time agents spend reading skill files and figuring out the exact commands to run. Skill files are essentially lossy compression of day-to-day workflows. When invoked in Codex or Claude, the agent spend reasoning tokens “decompressing” them back into actions, adding latency and cost. What if we traded a bit more upfront compression for nearly token-free decompression—turning user workflow trajectories directly into executable programs? Check out our early exploration by @demisama_.

Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on on Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 API: 🔗 Tech blog:

An annoyingly common question I get as an AI PhD student is “When can you get the ChatGPT AI to do something useful? It can’t even work on my phone yet. Siri is pretty dumb.” To be fair, I think their criticism is correct. I personally wished that AI was better integrated on my phone. LLMs can solve IMO problems, so shouldn’t it be a cakewalk for it to remind me of the text I forgot to respond to last week? Obviously not, since it doesn’t exist in my pocket yet. Or maybe Apple’s new update yesterday fixed this and my research project is obsolete. We are releasing iOSWorld ( a dynamic iPhone benchmark with 26 newly created apps grounded in personal context. Each of the 26 apps is centrally seeded around one persona, Jordan Avery, and the apps interact together in a realistic ecosystem that reflect real app interactions. We create 133 personalized mobile agent tasks to test in this environment, and the best model, even with privileged information, only scores 51%.

Congrats to the Webwright team at @MSFTResearch for taking the #1# spot on Odysseys, a highly challenging benchmark for long-horizon web agents: Odysseys evaluates realistic, multi-hour web workflows that require sustained planning, memory, reasoning, and verification across many websites and tools. These are far beyond short single-step browser tasks. For example, if you are searching for CS faculty positions, a single task could involve building a comprehensive Excel tracker of openings across the top CS schools using CSRankings as the master checklist; verifying every school directly through department, engineering, and university careers pages for CS/AI/ML/data science/robotics/vision faculty roles; opening and validating each posting; maintaining structured evidence and verification tabs; and finishing with a completeness audit and summary of hiring trends. Exciting progress toward truly capable long-horizon web agents.