Nous Hermes
Nous Research's steerable open-weight models, now Hermes 4 and 4.3
Overview
Nous Hermes is the open-weight language model family from Nous Research. Hermes 4 (14B, 70B and 405B) and Hermes 4.3 (36B, trained on the Psyche network) are hybrid reasoning models tuned to follow the user's system prompt with few refusals, published on Hugging Face under Apache 2.0 or Meta's Llama 3 licence depending on the model.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Nous Hermes
- A chat assistant that follows your own system prompt, with far fewer refusals than mainstream models
- Switchable step-by-step reasoning from the Hermes 4 hybrid mode
- Function calling and schema-valid JSON output from an open-weight model
- A private local assistant from Hermes 4 14B or quantised Hermes 4.3 36B
- Long-document work with Hermes 4.3's 512K-token context
- Hosted Hermes 4 405B through OpenRouter without provisioning your own GPUs
Hermes remains the open-weight line to reach for when you want a model that follows your own system prompt: Hermes 4 brought hybrid reasoning in 14B, 70B and 405B sizes, and Hermes 4.3 squeezes similar quality into 36B with a 512K context, small enough for a single workstation GPU when quantised.
Check the licence on each card, because the larger models carry Meta's Llama terms, and expect no newer Hermes model soon, since Nous's attention has moved to Hermes Agent. Hosted access now runs mainly through third parties such as OpenRouter.
What we found
Hermes is the family of open-weight language models from Nous Research, Inc., a Delaware company that began as a loose group of open-source fine-tuners. The first Nous-Hermes, a fine-tune of Llama 13B, was posted to Hugging Face in June 2023. Later versions sat on Mistral, Mixtral, Yi and Llama 3, and Hermes 2 Pro was built around function calling and JSON output. The idea has stayed the same: a model that follows the user's system prompt rather than a vendor's policy layer, with far fewer refusals than mainstream assistants.
The current generation is Hermes 4, released on 26 August 2025 in 14B, 70B and 405B sizes with a hybrid mode that thinks step by step only when asked. Hermes 4.3, released on 3 December 2025, is a 36B model on ByteDance's Seed-OSS base with a 512K context, which Nous says nearly matches Hermes 4 70B at half the size. It was the first Hermes trained on Psyche, Nous's distributed training network coordinated on the Solana blockchain. Nous calls it state of the art on RefusalBench, a claim it makes about its own model. No new Hermes language model has followed; since then Nous has released NousCoder-14B and, in February 2026, Hermes Agent, which has its own entry here.
Licences differ by model, so check each card. On Hugging Face, Hermes 4.3 36B and Hermes 4 14B are Apache 2.0, while the 70B and 405B carry Meta's Llama 3 licence. The weights are free. Hosted use is less obvious: on 14 October 2026 the Nous Portal catalogue listed no Hermes model, while OpenRouter offered Hermes 4 405B at $1 per million input tokens and $3 per million output.
Money comes from crypto and mainstream investors. Fortune reported a $50 million Series A led by Paradigm in April 2025 at a $1 billion valuation for an unlaunched token, and in October 2026 Nous announced $90 million from investors including NVIDIA, Microsoft's M12 and Samsung. For teams wanting steerable weights they can host, Hermes remains a strong choice.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Nous Research is a fully decentralised research collective with no official company registration — entirely volunteer contributors working open source.
- Hermes 3 achieved performance comparable to Llama 3.1 8B Instruct while having fewer refusals on legitimate edge-case questions.
- The Nous Research community developed the synthetic training data technique of using GPT-4 to generate diverse instruction datasets, which became a standard approach across the open-source model fine-tuning community.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Nous Hermes's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Nous Hermes unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Nous Hermes directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to Nous Hermes
More in LLMs & Chat
