Nous Research

NousCoder 14B

Nous Research

NousCoder 14B is Nous Research’s independently released open-weight language model. This competitive programming specialization uses reinforcement learning on 24,000 verifiable coding problems. Training ran on 48 B200 GPUs for four days. The model card reports LiveCodeBench v6 improvements under a fixed evaluation period; it is not an evaluation of full repository maintenance. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.

Nomos 1

Nous Research

Nomos 1 is Nous Research’s independently released open-weight language model. Developed with Hillclimb AI, this mathematical specialization generates natural-language solutions and proofs. It is designed for the separately released Nomos reasoning harness. Its Putnam comparison measures the model plus that harness, rather than single-draw model-only accuracy. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.

Hermes 4.3

Nous Research

Hermes 4.3 is Nous Research’s independently released open-weight language model. Based on ByteDance’s Seed-OSS 36B Base, this generation is the first production Hermes checkpoint post-trained through the distributed Psyche network. An extended context and larger training run strengthen general responses and reasoning. A centrally trained comparison is a research variant, not a separate product page. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.

Hermes 4

Nous Research

Hermes 4 is Nous Research’s independently released open-weight language model. This hybrid reasoning family uses a substantially expanded post-training corpus with verified reasoning traces. It supports both ordinary responses and explicit thinking for mathematics, code, STEM, and creative tasks. The smaller Qwen-backed 14B and larger Llama-backed models have different base-model licenses. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.

DeepHermes 3 Preview

Nous Research

DeepHermes 3 Preview is Nous Research’s independently released open-weight language model. The first named DeepHermes preview combines ordinary chat with controllable long-chain reasoning. It is a distinct reasoning post-training, rather than a generic inference setting on an unchanged Hermes model. The card recommends particular prompting and sampling settings for reasoning mode. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.

Hermes 3

Nous Research

Hermes 3 is Nous Research’s independently released open-weight language model. The published August 14 technical report establishes the family’s research release; the official catalog dates the 405B model to August 24 and the 3B extension to December 11. It improves function calling, roleplay, conversation coherence, and structured responses. Downloadable weights enable private deployment and further research. This page preserves checkpoint and evaluation scope; tool execution requires an external harness, and generated answers or proofs need independent verification.