Oleksii Kuchaiev

1/4 We see no wall in post-training. Scaling RL software, infra, and data keeps yielding major capability gains. We trained across 30 RL environments with up to 4,000 instances per batch — math, code, STEM, agentic tool use, SWE, terminal, safety — all in a unified multi-environment RLVR setup.

Nvidia just crossed Google as the biggest org on @huggingface with 3,881 team members on the hub. I'm officially calling it: Nvidia is the new American king of open-source AI!

Nemotron 3 Nano and Super from @nvidia are now available on Tinker! We're offering a limited-time GTC 50% discount for both. The Nemotron family features open hybrid MoE models optimized for compute efficiency for agentic applications.

Breaking: Russia is secretly providing satellite imagery and sharing drone technology to help Iran target U.S. forces in the region on.wsj.com/478OhaN

Shoutout Joe Nemotron (@llm_wizard)

NeMo just grew some claws

If you are attending GTC 2026, stop by tomorrow, Tuesday, March 17  at 3:00 p.m. - 3:40 p.m. PDT to learn about how we post-train Nemotrons. Looking forward to seeing everyone! Session number: S81558

An angry suicidal teenager is not going to be able to bypass frontier models' gurdrails. And even if he does, no lab is going to synthesize it or even ship components to the same place. A bans on this exists since at least 1975 (see Asilomar Conference banning certain recombinant DNA experiments).

The question here is: Can an angry, suicidal teenager imagine a doomsday virus?

@rasbt This is beautiful

I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place! sebastianraschka.com/llm-architectu…

@natolambert Adoption-wise, I agree, closed will be far ahead, especially on a consumer side. But in terms of capabilities, I do not think the gap will persist for long. Will be very much like closed OS (iOS, Mac, Windows for consumers) vs Linux/Unix (which rules the Internet).

World will converge on 3 types of models 1. Closed frontier (Ant, OAI, Gemini) 2. Open frontier (2-3 labs, much consolidation coming) 3. Open small / tool (fairly empty now) The open frontier will be far from the closed frontier, but way cheaper. Other statements are cope.

NVIDIA’s Nemotron 3 Super is now available in Perplexity, Agent API, and Computer.

We collaborated with @NVIDIA to teach you about Reinforcement Learning and RL environments. Learn: • Why RL environments matter + how to build them • When RL is better than SFT • GRPO and RL best practices • How verifiable rewards and RLVR work Blog: unsloth.ai/blog/rl-enviro…

@SenAdamSchiff whenever putin needs something, trump quickly delivers

There are few people in the world who have derived more benefit from our war with Iran than the Russian dictator, Vladimir Putin. We explain why:

We 💚 open models

🦞These innovations come together to create a model that is well suited for long-running autonomous agents. On PinchBench—a benchmark for evaluating LLMs as @OpenClaw coding agents—Nemotron 3 Super scores 85.6% across the full test suite, making it the best open model in its class.

4/4. I'll go deep on the post-training methodology at GTC 2026 Conference — session S81558, Tuesday March 17, San Jose. Will cover RL scaling across environments, what worked in agentic training, and where we think this is headed.

3/4. The entire post-training stack is open (Apache 2.0): NeMo-Gym (env collection and orchestration): github.com/NVIDIA-NeMo/Gym NeMo-RL (RL training): github.com/NVIDIA-NeMo/RL All SFT + RL data also open on HuggingFace. PRs and new environments welcome! Full tech report: research.nvidia.com/labs/nemotron/…

Nemotron 3 Super is here — 120B total / 12B active, Hybrid SSM Latent MoE, designed for Blackwell. Truly open: permissive license, open data, open training infra. See analysis on @ArtificialAnlys Details in thread 🧵below: