Nous Research

Introducing NousCoder-14b, a competitive olympiad programming model. Our latest blog details the full findings from extensive experiments and logs with the full stack released - the RL environment, benchmark, and harness built in Atropos, all fully reproducible with our open training stack. NousCoder-14b was post-trained on Qwen3-14B by researcher in residence Joe Li using 48 B200s over the course of 4 days, our Atropos framework, and Modal's autoscaler. It achieves a Pass@1 accuracy of 67.87%,+7.08% over Qwen's baseline accuracy using verifiable execution rewards. View the full blog post here: https://lnkd.in/gAQc9rmk

Today we open source Nomos 1. At just 30B parameters, it scores 87/120 on this year’s Putnam, one of the world’s most prestigious math competitions. This score would rank #2/3988 in 2024 and marks our first step with hillclimb towards creating a SOTA AI mathematician. Nomos 1 achieved an 87/120 with 8 perfect scores, while Qwen3-30ba3b-Thinking-2507 scored 24/120 when run in the same harness under the same conditions, indicating that the performance is largely due to post-training and data quality rather than the harness. Submissions were blind-graded by a human Putnam top 200 contestant who was given anonymized submissions. The exact files sent to our human annotators for grading are available here de-anonymized here: https://lnkd.in/ejUeS8JC, along with the runbooks used to generate them https://lnkd.in/exNc4-kT omos/tree/main/runbooks. We open-source our model here https://lnkd.in/eHvQfMZK and our reasoning harness here https://lnkd.in/e32B9aeK.

We are excited to announce a drastic 70%+ price reduction on our flagship hybrid reasoning models. Try them now at https://lnkd.in/e6g5NF8z

Check out Vercel's blog for a little article on how we use use BotID at Nous Research: https://lnkd.in/g9gWwdT5 . Thanks Vercel folks, you're always a pleasure to work with. :)

We are excited to announce our partnership with hillclimb to create a model that can take on the world's most difficult mathematics challenges.

As large language models continue to scale, centralized pre-training faces growing challenges in high barrier to entry, data availability, compute concentration, and trust. Decentralized pre-training offers a promising alternative by leveraging globally distributed compute and data while maintaining model quality and security. Watch to the full talk from the PyTorch conference by our Chief Scientist, Bowen Peng. https://lnkd.in/eEjTCQZn

This has been a fun release to work on, because it pulled together many threads we've been juggling for a while: - The Hermes 4 models themselves (which are genuinely distinctive imo) - The custom RL framework that made fine-tuning possible - A fresh and rather experimental chat UI (won't be to everyone's taste, but no point in us all building the same chat over and over!) - Lots of behind the scenes tweaks to our developer portal for key management, credit tracking, and payments, etc.. - An auth provider migration (to Privy.io) - Loads of explorations and comparisons of different inference providers - An innovative cross-chat memory system, backed by a fancy a knowledge graph Lots of moving parts that clicked into place. Nice work team. :)

Nous Research presents Hermes 4, our latest line of open-source models. https://lnkd.in/eFf3WCbG Hermes 4 builds on our legacy of user-aligned models with expanded test-time compute capabilities. Special attention was given to making the models creative and interesting to interact with, unencumbered by censorship, and neutrally aligned while maintaining state of the art level math, coding, and reasoning performance for open weight models. You can try Hermes 4 in the new, revamped Nous Chat UI. https://lnkd.in/ewQ6Mj4P For the first week, all Hermes 4 inference in Nous Chat is free of charge.

Measuring Thinking Efficiency in Reasoning Models: The Missing Benchmark https://lnkd.in/eJFWZRsb We measured token usage across reasoning models: open models output 1.5-4x more tokens than closed models on identical tasks, but with huge variance depending on task type (up to 10x on simple questions). This hidden cost often negates per-token pricing advantages. Token efficiency should become a primary target alongside accuracy benchmarks, especially considering non-reasoning use cases. Read the thorough review of reasoning efficiency across the open and closed model landscape in our latest blog post in collaboration with our researcher in residence, Tim. See more of their work here: https://lnkd.in/eju3abfC

Announcing the Nous RL Environments Hackathon in SF! Create with Atropos, Nous' RL environments framework, and claim your stake of a $50,000 prize pool. Partners include xAI, Nvidia, Nebius, Akash Network, Lambda, TensorStax, Runpod and Cerebral Valley. You can also use the ⁠atropos-rl-gym channel in our Discord to discuss and learn in preparation!   May 18th. Sign up below 👇👇 https://lnkd.in/gjusg2x9 https://lnkd.in/eFzRqbXa