Nikhil the Coder @NikhilCoderAI

It always amazed me how popular OCR models are in the OSS community, and we collaborated together with the Baidu team to bring their latest SOTA OCR model in-tree into vLLM! Check it out!

vLLM @vllm_project

14 hours ago

🎉 Unlimited-OCR from @Baidu_Inc now runs in vLLM. One-shot parsing of entire books with constant KV cache, powered by Reference Sliding Window Attention (R-SWA). 🧠 R-SWA keeps KV cache fixed throughout decoding — no memory blowup, no slowdown, no matter how long the output

7 61 657 44K 489

the fact that i dont have access to gpt 5.6 sol because of some bullshit is very very frustrating.

Really proud to see Nemotron 3 Ultra generalizing to agentic tasks that didn't exist when we trained it!

NVIDIA AI @NVIDIAAI

2 days ago

.@ArtificialAnlys just dropped a brand new leaderboard called AA-Briefcase for evaluating realistic tasks in complex projects. Nemotron 3 Ultra ranks among the top open models, with strong performance across a wide range of long-running agentic tasks, even when encountering them

27 43 277 56K 37

The nsa just gave china a win

Matthew Berman @MatthewBerman

2 days ago

Anthropic just struck a deal with the government to allow 100 select companies and governmental agencies use Mythos. The government and anthropic are now deciding who uses frontier intelligence. Hopefully this is just Mythos and not the standard for all frontier models going

150 54 775 206K 94

What's with the OpenAI codex use limits? I have both ChatGPT Business and Plus subscriptions. I can now barely get 1500-2000 lines of C++ code with per week. And 5 h limits are gone in <30 minutes. They are forcing me to other solutions (Grok build & opencode with Kimi/DS 4 Pro)

I don't buy the safety story. Not from the labs, not from the government. Fable, gone. Mythos, gone. Then all of GPT-5.6, down to Luna, the cheap everyday model for summarizing and drafting. A summarizer is not a cyber weapon. But cheap intelligence is the one thing that lets everyone else catch up. So they locked it down too. It was never about security. It's about who gets the frontier, and who stays behind.

Anthropic @AnthropicAI

2 days ago

Since June 12, we’ve been working closely with the US government to restore access to Claude Mythos 5 and Fable 5. Today, the government notified us that Mythos 5, our strongest cybersecurity model, can be redeployed to a set of US organizations that operate and defend critical

2K 3K 30K 4.6M 3K

NVIDIA just released Fast-FoundationStereo on Hugging Face A real-time foundation model for zero-shot stereo depth estimation 10× faster than FoundationStereo with matching zero-shot accuracy bringing instant 3D perception to robots and devices.

Within a month, we got both Agent's Last Exam and OSWorld V2. Agent benchmarks are starting to look very different. They now cover workflows and domain expertise that few people would have imagined benchmarking before, spanning real professions, industries, and long-horizon tasks instead of toy problems. I think these benchmarks will increasingly shape the direction of foundation model training over the next few years. They are becoming optimization targets, not just evaluation suites.

XLANG NLP Lab @XLangNLP

2 days ago

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved? Not even close. 🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks. What's new:

13 103 396 131K 170

1/N Neuroscience and social science research on humans has shown: – Similar brain activity predicts friendship and cooperation – Diverse minds drive innovation We wondered whether AI-AI interaction would show the same pattern. It does. LLMs with similar internal representations cooperate more, but produce less novel output. 🧵 (ICML 2026)

Key results: • Training Optimization On GPT nanochat training optimization, AutoScientists reaches the same validation performance in just 34 experiments compared to 65 for AutoResearch, delivering a 1.9× improvement in search efficiency. Starting from a stronger baseline where the single-agent system stalls, AutoScientists finds 7 successful improvements in 93 experiments, while AutoResearch finds none across 100 attempts. • Biomedical Machine Learning On BioML-Bench, which covers 24 tasks across medical imaging, drug discovery, protein engineering, and single-cell omics, AutoScientists achieves a mean leaderboard percentile of 74.4%. This surpasses the strongest previous biomedical AI research agent by 8.3 percentage points and successfully completes every task in the benchmark. • Protein Engineering For ProteinGym supervised fitness prediction, AutoScientists discovers an extension to Kermut on the ACE2-Spike dataset, improving Spearman correlation from 0.747 to 0.840. The same approach transfers across all 217 ProteinGym assays, raising the benchmark average from 0.657 to 0.700 and establishing a new state-of-the-art result on the supervised substitution benchmark. The interesting part is not just the performance gains. It is the research process itself. AutoScientists shows how decentralized AI agents can organize around evidence, learn from failures, and continuously reshape their search strategy, much closer to how real scientific teams operate.

No organization operates in isolation anymore. Businesses depend on customers, suppliers, partners, regulators, financial institutions, and technology providers that all rely on different systems. The ability to exchange information seamlessly is as important as the information itself. Organizations that build products which integrate easily with existing workflows reduce adoption barriers, shorten implementation timelines, and create better customer experiences.

As #ICML2026 is excitingly near, here are some of the most-cited papers and most-starred git repos: Code and workflows to map papers to Semantic Scholar citation data and github repos is made publicly available.

keyboard_arrow_left Previous keyboard_arrow_right Next

what i like here, scaffold is a learnable object that co-evolves with the policy, not a fixed harness. but only the inner policy scaffold mutates (memory, error handling, orchestration), outer trust boundary stays frozen. that boundary split is the anti-reward-hacking trick, you can't touch env/tests/tools, just the orchestration logic around them.

Ornith @ornith_

3 days ago

Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding. Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on

480 997 7K 5.2M 6K

🚀 Day-0 support for LFM2.5-230M on vLLM! Ready to serve @liquidai's smallest LFM2 model for fast GPU inference and agentic workloads. 🧠 230M parameters, built on the LFM2 architecture 📚 Pre-trained on 19T tokens with 32K context 🛠️ Designed for agentic tasks across phones, robots, home automation, and network automation devices ⚡ Built to run across CPUs, NPUs, and GPUs Congrats to the @liquidai team on the launch! Recipe: recipes.vllm.ai/LiquidAI/LFM2.…

Liquid AI @liquidai

3 days ago

Introducing LFM2.5-230M: our smallest model yet, built to run fast anywhere (CPUs, NPUs, and GPUs) to enable agentic tasks on phones, robots, home and network automation devices. > 230M parameters, built on the LFM2 architecture > Pre-trained on 19T tokens, with a 32K context

81 195 2K 230K 1K

Congrats! Open source GLM model is really a game changer! Extremely fast, cheap, and high quality!

Dmytro Dzhulgakov @dzhulgakov

4 days ago

you may have heard that glm-5.2 at 280 token/s is cool, how about 318 and we still have room to go

55 29 659 262K 173

@brianringley "It's really not enough to have the world's best hardware, which I believe we do..." "It has to be reliable." "It has to be serviceable." "It has to be cheap and affordable." "All of those things are what will actually make for a successful humanoid product in the long run."

Recruiting is not something you do just 9-to-5, I worked hard on this one

Natalie Biderman @Natalie_Bid

5 days ago

I've spent years studying how human memory works: how we learn and forget, and how our memories shape what we do. I'm thrilled to share that I've joined the founding team of Engram, which is coming out of stealth today. Now is the time to understand how to build machines with

2 8 71 27K 37

Someone at Google just shared a 19-page PDF that completely changed how I think about building AI agents. It introduces a simple but powerful loop: → Act → Observe → Learn → Repeat The LLM makes a change. A real system tests it. The results come back. The model learns from that feedback and tries again. No extra training. No fine-tuning. Just continuous improvement driven by real-world feedback inside the context window. Want an agent that actually gets better while working? This PDF is worth your time. Read it first, then explore the article below.

divyansh tiwari @DivyanshT91162

a week ago

x.com/i/article/2068…

6 11 22 10K 7

OpenAI engineer: “90% of our engineers use Codex subagent harnesses + loops to ship code 5× faster. close the agent loop. Give an agent a way to verify its own output.” in 30-minute talk, an OpenAI engineer explains how to build self-improving agentic systems with harnesses. Worth more than a $500 agentic course online. Watch it today, then read how hedge funds apply loop engineering in trading.

divyansh tiwari @DivyanshT91162

5 days ago

x.com/i/article/2069…

13 16 38 6K 39

Harnesses seem to soon matter more than the model as many models are already good enough. That is why I'm building from ground up my own harness / the Swiss army knife of AI called "pkchat" using C++ and as few external libraries as possible. It has no agentic mode yet, but it already has many features: interactive chat, command line tools, benchmark mode (I'm still busy making my own internal benchmark set, will be rather small max 100 tasks) and (very simple) text editor. My design philosophy: 1. Speed 2. Robustness e.g. Lots of internal unit tests, integration tests, mockup server etc. Of course using Valgrind to verify it has no memory leaks. 3. Ease of use. 4. Lots of built-in features (less need to call external tools, which is both potential security risk, but also can slow things down in agentic use) 5. Modular design and well-planned roadmap from the beginning (2k+ lines in PLANS.md alone, plus TODO and AGENTS.md) UI and the features have a lot of rough edges and I think I maybe should just release even though it will disappoint many, who expect a polished product (it is not). Things like low memory use aren't top in the top criteria, but so far it starts < 1 sec and uses less than 1 MB and optimized executable also under 1 MB, so I think it will likely work well on very low memory devices. I know everybody else is pretty much using TypeScript + JS frameworks or Python and their harnesses are getting huge and often slow. I want things to complete and start in an eye blink.