Hugging Face

New live stream on training agents. Session 3 is on GRPO. Sessions 1 and 2 taught the model to imitate, and now we'll train on outcomes/results. This is a practical session, with code-along tutorials and reproducible experiments to explore. We’ll also answer questions as we go. What is GRPO? In short, you sample a group of completions, score them with a reward function, and use relative scores as the learning signal. There's no reward model or critic, just simple verifiers and reward functions. We'll also dive into the failure modes: reward hacking, collapse, and length gaming. https://lnkd.in/eskatUdE

Training Agents, Session 3: from a fine-tuned baseline to reinforcement learning with GRPO. In this live session, we'll move past imitation and start training on outcomes. We'll take the model from the earlier sessions and run Group Relative Policy Optimization: sample multiple completions per prompt, score them with a reward function, and use the group's relative scores as the training signal — no value model, no reward model, no critic to train. Everything runs in TRL. What we'll cover: - Why GRPO after SFT, and what RL adds that imitation can't - How GRPO works: group sampling, relative advantages, and why that replaces a critic - Writing verifiable reward functions for agent tasks - Running TRL GRPO on Hugging Face Jobs - Reading the training curves: reward, KL, and completion length - Common failure modes: reward hacking, collapse, and length gaming Repo: https://lnkd.in/e7CZsTjZ This is part of the Training Agents series: using coding agents to design, run, monitor, and review post-training experiments, while training models to become better agents.

Today we’re releasing Laguna S 2.1, our new open-weight model for agentic coding and long-horizon work. Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token, up to 1M tokens of context, and thinking and no-thinking modes. It is capable enough to compete with models several times its size, yet small enough to run locally on a single NVIDIA DGX Spark. What sets Laguna S 2.1 apart is its persistence. Across long-horizon coding and research tasks, it holds onto a goal, uses tools, checks its work, recovers when an approach fails, and continues making progress for hours with little or no intervention. That persistence comes with a practical balance of cost, speed, and ownership. Laguna S 2.1 weight class is designed to take on real, long-running agentic work at a cost and speed that make it practical to run often and at scale. Because it activates only 8B parameters per token, long agent runs and reinforcement learning loops are faster and less expensive than they would be with much larger models. We’re releasing Laguna S 2.1 under OpenMDW-1.1 with checkpoints in BF16, FP8, INT4, and NVFP4, alongside official GGUF and MLX quantizations. The weights are available today on Hugging Face. Run it through pool, vLLM, SGLang, Ollama, llama.cpp, ZML, MLX, or NVIDIA TensorRT-LLM, or access it through OpenRouter and the Poolside API.

Join us this Tuesday to learn more on Local AI, from software to hardware 🤗 We'll be joined by Ahmad M. Osman & Michael Bradley covering hardware setups & local inference with live demo, and Alex Cheema & Sherif Cherfa on picking model for your hardware, model compression and REAPs 🔥 Set your reminders to not miss out! 🔔

Making Meta's "Recycling the Web" dataset likely cost ~$369k in compute in 2025. Today, the same data would cost around $28k. I checked for $11! The recipe (COLM 2025) rewrites discarded, low-quality web pages from Common Crawl Foundation into pre-training data with an LLM. Meta used Llama-3.3-70B as the rewriting teacher. I got a coding agent to re-implement the pipeline with datatrove on Hugging Face Jobs. The result: in 2026, a 9B model beats the paper's own 70B teacher. - Qwen3.5-9B clears the quality gate 61.7% of the time vs the 70B's 49.6% - Blind paired judges prefer the 9B's rewrites 20-4 - The whole PoC: ~$11 in GPU time, 4 models compared - The winning 9B arm: $0.60 for 600 documents — about $0.64 per million kept tokens - At that rate, Meta's released 44.4B-token set would be ≈$28k to reproduce; at the 70B teacher's rate ($8.30/M), ≈$369k (naive linear extrapolation) The original release: the quality filter was never published, and only part of the data shipped. This version is fully open i.e. the datatrove pipeline, all 1,953 rewrites, the cost receipts, and a 30-second CPU script that re-scores every rewrite. If you don't trust my agent's eval, run the check yourself! REPRODUCE.md walks the whole pipeline; scaling up is two numbers in the executor. If you rebuild the full 44.4B-token set, I'd love to see it! https://lnkd.in/em4huThV

If you're looking to have a load of fun with agents and learn about cutting-edge AI, try this reproduction challenge. You'll grab a snippet from the space, drop it in in your preferred agent (codex, claude, pi etc), and pick a paper to reproduce from this year's ICML. Along the way, your agent will share its status, any artifacts, and learnings in a logbook. My favorite bit is that you can just chat with the context of your agent or others' if you want learn about papers. https://lnkd.in/ezQT4byW

NVIDIA AI officially shipped an integration with Diffusers in their widely popular NeMo AutoModel library! This means the community can load and export any model from Diffusers and use AutoModel to meet their fine-tuning and pre-training needs. This integration is also not just another training library. It carefully balances the needs between accessibility and seamless scalability. More specifically, one can take advantage of memory-efficient sharding, latent caching, multiresolution bucketing, and configurations that scale gracefully from one GPU to hundreds of GPUs. I had the honor of closely collaborating with the NVIDIA team (Pranav Thombre, Linnan Wang, Ph.D., and Alexandros) on this. We have published a blog post to tell you all about it. Please check out the link in the first comment.

this week we had everything: > PrismML new 27B ternary & binary models 🪴 10x smaller with 95% equally smart to fp16 🥵 https://lnkd.in/e5wcvAju > Thinking Machines Lab Inkling 🖋️ open sota multimodal reasoning Omni > Kimi (Moonshot AI) 3T sota K3 going open-source announcement > TODAY! Qwen announcing a new rivalling 2.4T open model announcement read our architecture unpack and vibe-test of Inkling here, find model repository there https://lnkd.in/eqYM_pBW follow Kimi and Qwen here to stay tuned! https://lnkd.in/e3Z8k5Si https://lnkd.in/evAPuFJc