Nous Research releases Hermes 3

Nous Research released Hermes 3, a family of open-weight fine-tunes of Meta’s Llama 3.1 at 8B, 70B and 405B parameters, trained on a largely synthetically generated dataset built to strengthen instruction-following and adherence to system prompts.

The release emphasised capabilities aimed at developers building agents and character-driven applications rather than raw benchmark leadership: multi-turn conversation, long-context coherence, complex roleplay, an internal reasoning mode, and improved function-calling for tool use. Nous framed the project around what it called “individual alignment” — steering the model to follow a user’s own instructions and persona closely rather than a single fixed set of default behaviours — a positioning that distinguished it from the safety-tuned defaults of closed frontier assistants. The company reported performance comparable to or exceeding the base Llama 3.1 models on its own evaluations.

Releasing a full range from 8B up to the frontier-scale 405B variant meant Hermes 3 was one of the largest fully open fine-tunes available at the time, extending Nous Research’s series of Llama derivatives (Hermes and Hermes 2 had covered earlier Llama generations) into Meta’s newest release within weeks of its own July 2024 launch. The 405B model, hosted with infrastructure partner Lambda, drew additional attention after VentureBeat reported users triggering unusual, seemingly unintended “existential crisis” style outputs from the model under certain prompts — an example of the harder-to-predict behaviour that comes with releasing full weights for a model of that size rather than a filtered API.