Skip to content

hermes.md

Latest commit

 

History

History
41 lines (28 loc) · 1.5 KB

hermes.md

File metadata and controls

41 lines (28 loc) · 1.5 KB

Hermes (Nous Research agent framework)

Hermes is Nous Research's open agent framework. It works with any OpenAI-compatible local provider.

Disambiguation: this page is about the Hermes agent framework, not the Nous-Hermes family of LLMs (those would be the "model"). afm also supports the hermes tool-call format used by Hermes-derived models, auto-detected when model_type matches.

1. Start the server

afm mlx -m mlx-community/Qwen3-Coder-Next-4bit \
  --port 9999 --enable-prefix-caching --concurrent 4

2. Configure Hermes

Set the OpenAI base URL to afm:

export OPENAI_API_BASE=http://localhost:9999/v1
export OPENAI_API_KEY=x

Or in Hermes config (e.g. YAML/JSON depending on your setup):

provider:
  type: openai
  base_url: http://localhost:9999/v1
  api_key: x
  model: mlx-community/Qwen3-Coder-Next-4bit

3. Run Hermes

Hermes will issue chat completions with tool calling. afm's auto-detected tool format and harmony/think reasoning extraction work with Hermes' loop.

Tips

  • For Hermes' self-improvement loops (trajectory generation), enable structured output: pass response_format: {"type":"json_schema","json_schema":{...,"strict":true}} and turn on --enable-grammar-constraints server-side.
  • Multi-step trajectories benefit from --enable-prefix-caching.
  • Deterministic seeds: pass seed: <int> per request — afm threads it through end-to-end (logged via --vv for reproducibility).