fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_mlx

gpt-oss-20b · Hermes-Agent tool finetune · MLX

Apple Silicon native. Runs on M-series Macs through MLX with no PyTorch detour. Tested on M2 Max and M3 Pro.

• Format — MLX (safetensors + index)
• Size on disk — ~12 GB
• Unified memory needed — 24 GB minimum, 32 GB comfortable
• Recommended runtime — mlx-lm ≥ 0.18

What this is

A tool-use finetune of OpenAI's gpt-oss-20b for Hermes-Agent, a local agent framework that needs models which call tools reliably, follow multi-turn instructions, and don't argue with system prompts.

The base model is the 21B-parameter (3.6B active) Mixture-of-Experts release from OpenAI. This finetune preserves the Harmony chat template and the reasoning-effort knob, and improves:

• Function-calling adherence (correct JSON, no commentary mid-call)
• Long agent loops (10+ turns of tool → observe → plan)
• System-prompt fidelity (respects role boundaries and refusal/allow-list rules)

It is not affiliated with NousResearch's Hermes model series. "Hermes-Agent" here refers to the local agent framework only.

Quickstart

pip install -U mlx-lm

One-shot generate

mlx_lm.generate \
  --model fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_mlx \
  --prompt "List three bash one-liners that find files larger than 100 MB." \
  --max-tokens 256

Local OpenAI-compatible server

mlx_lm.server --model fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_mlx --port 1234

Point Hermes-Agent (or any OpenAI client) at http://127.0.0.1:1234/v1.

Hermes-Agent integration

Add a profile in ~/.hermes/config.yaml:

profiles:
  gpt-oss-20b-tools:
    provider: openai
    base_url: http://127.0.0.1:1234/v1   # LM Studio / vLLM / mlx_lm.server
    model: fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_mlx
    temperature: 0.7
    top_p: 0.95
    min_p: 0.1                            # important for MoE stability
    max_tokens: 8192
    tool_choice: auto

Then hermes profile use gpt-oss-20b-tools and the agent loop will route tool calls through this model.

Sampling

Param	Value	Why
temperature	0.7	balanced; drop to 0.2 for strict tool calls
top_p	0.95	standard nucleus
min_p	0.1	required for MoE — prevents dead-expert tokens
repetition_penalty	1.0	the model handles repetition itself

Harmony reasoning effort: set the system message to Reasoning: low|medium|high. high is roughly 3-4x more output tokens but noticeably better on multi-step tool plans.

Training

• Base: openai/gpt-oss-20b
• Method: LoRA SFT (rank 64, alpha 16) merged back into BF16
• Frame: Unsloth + TRL on a single H100 (80 GB)
• Data: ~42k tool-use traces from Hermes-Agent sessions, filtered for successful tool calls and clean JSON. No synthetic distillation.
• Length: 8192 tokens, packing on
• Loss: assistant-only, mask user/system/tool

The _16bit repo holds the merged BF16 weights. The _4bit, _mlx, and _gguf repos are quantizations of that checkpoint.

Limitations

• Math and code-generation are unchanged from the base — this finetune optimizes the agent loop, not raw reasoning.
• The model can over-call tools when given vague instructions. Add a "if you can answer directly, do so" line to the system prompt.
• English only. Other languages were not in the training mix.
• Not safety-tuned beyond what gpt-oss-20b already provides.

Other formats

• BF16 reference — full precision, vLLM / Transformers
• MXFP4 4-bit — fits a 16 GB GPU
• MLX — Apple Silicon native
• GGUF — llama.cpp / Ollama / LM Studio

License

Apache-2.0, inherited from the base model. No additional restrictions.

Citation

@misc{fesalfayed_gptoss20b_hermesagent_2025,
  author = {Fayed, Fesal},
  title  = {gpt-oss-20b Hermes-Agent tool finetune (mlx)},
  year   = {2025},
  url    = {https://huggingface.co/fesalfayed/gpt-oss-20b-hermes_agent-tool-finetune_mlx},
}