Hermes 4

About Hermes 4

Hermes 4 is a family of open-weight language models from Nous Research, released in August 2025 as the successor to Hermes 3. The models are post-trained on a large synthetic corpus of reasoning and non-reasoning data and are designed to be steered by the user's system prompt rather than by a fixed vendor alignment. Nous Research evaluates the family with its own RefusalBench benchmark, which measures how often models decline requests in categories that mainstream assistants commonly restrict.

The 14B version is based on Qwen3-14B and is licensed under Apache-2.0; the 70B and 405B versions are based on Llama 3.1 and use the Llama 3 license. Hermes 4 has a hybrid mode that can produce explicit thinking steps before answering, and it supports tool calling and structured JSON output. Community GGUF quantizations of the 14B model range from about 7 GB to 16 GB, which suits local runtimes such as Ollama, LM Studio and KoboldCpp on consumer GPUs.