We have just released v0.2.0 of Atropos - Nous' RL environments project! New environments, updated handling for api endpoints that don't support n iterations from the openai spec, added better TRL support, and an official Trainer partner, Axolotl! Read the changelog here:

So many updates and the hackathon hasnt even begun! You can now train using Atropos using Axolotl too

🚀 Async-RL is now available with Nous Research Atropos and Axolotl! We've created a plugin for tighter RL iterations that leverages Atropos' optimized asynchronous rollouts. Check out the plugin and get started: github.com/axolotl-ai-clo…

Come hack with us!

Announcing the launch of Psyche nousresearch.com/nous-psyche/ Nous Research is democratizing the development of Artificial Intelligence. Today, we’re embarking on our greatest effort to date to make that mission a reality: The Psyche Network Psyche is a decentralized training

Largest distributed pretrain ever

Today marks a really big achievement for Nous, but also potentially the AI Landscape. We have begun a decentralized pretraining run of what is basically a dense Deepseek - 40B parameters, over 20T tokens, with MLA for long context efficiency. All checkpoints, unannealed,

Fun story about getting this to work -- since we're doing a dense model it was important to do tensor parallelism inside of MLA but we kept getting weird divergences DeepSeek skipped this since they had smaller experts that could fully fit within a single GPU (they detail this

excellent work by Jason Weston & team—extending our "Generative Reward Models" work with RL (GRPO) to optimize LLM reasoning during judgment scalable (synthetic) evaluation continues to be AI's key bottleneck!

There's like 5 or 10 slots left for the most promising potential competitors - Sign up now!

fate awaits at Nous Research hackathon opening 🔥

Nous Research's RL Environments Hackathon recap thread! Starting with the stars of the show, the winners! Top 3 for the subjective track were: 1st - Pokemon Trainer - by Ajay Uppili Arasanipalai & Alex Reibman 🖇️ 2nd - VR-CLImax by Jake Boggs 3rd - DynastAI by David van Vliet and

🧵For Claude Opus 4, we ran our first pre-launch model welfare assessment. To be clear, we don’t know if Claude has welfare. Or what welfare even is, exactly? 🫠 But, we think this could be important, so we gave it a go. And things got pretty wild…

Six teams just won $50,000 at Nous' first ever RL hackathon 🤩 Check out the winning demos👇 Nous Research xAI NVIDIA Nebius Akash Network Lambda TensorStax RunPod

Confused about recent LLM RL results where models improve without any ground-truth signal? We were too. Until we looked at the reported numbers of the Pre-RL models and realized they were serverely underreported across papers. We compiled discrepancies in a blog below🧵👇

Huge fan of Claude Code—so I built a python version using smolagents! Introducing SmolCC 🤖📟🛠️ An open source coding agent with Claude Code style tools (bash, grep, edit…✨) that can be easily customized.

Sequential Monte Carlo (SMC) is a powerful approximation method where multiple branches, or “particles”, are sampled, weighted, and resampled against a scoring function to produce likelier completions that fit the constraints.

absolutely ecstatic to announce my latest research with Nous Research in controlled text generation! this has been a very challenging and rewarding experience over the past ~6 months, I’m so happy to finally put it out into the world!