🜂 Atanor-4B
A fine-tune of Qwen3.5-4B specialized for agentic tool-use inside Hermes Agent.
There were 9B and 27B versions of Qwen fine-tuned to run as agents in Hermes (the Carnice models). That made me wonder: could a smaller model do the same job?
Atanor-4B is my answer — and my first fine-tune ever. It was trained entirely locally, on a single RTX 3090.
The name Atanor is the alchemist's furnace: the 4B is the lead that goes in, the agent is what comes out. 🜂
Results — agentic evaluation
Measured on a 60-task Hermes-native agent benchmark (real tool execution inside Hermes Agent, deterministic / temperature 0), base vs fine-tune:
| Metric | Qwen3.5-4B (base) | Atanor-4B |
|---|---|---|
| Agent score | 0.81 | 0.84 |
| Picking the right tool | 30% | 60% ⬆️ doubled |
| Task success | 67% | 73% |
The core agent skill — choosing the correct tool for a task — doubled (30% → 60%).
How it was made
Following the Carnice recipe, in two LoRA stages (BF16, on one RTX 3090):
- Stage A — reasoning repair: Bespoke-Stratos + NuminaMath-CoT
- Stage B — Hermes tool-use: the
kai-os/carnice-glm5-hermes-tracestraces (the full set, seq len 16384)
~33 hours of training, zero crashes.
Files in this repo
| File | What it is |
|---|---|
atanor-4b-full-Q4_K_M.gguf | Quantized (~2.6 GB) — run it directly in llama.cpp / Hermes / Ollama |
atanor-4b-full-Q4_K_M-f16.gguf | Full-precision GGUF (~8 GB) — for re-quantizing or lossless inference |
*.safetensors (merged) | Full merged model for transformers / further fine-tuning |
adapter/ | The LoRA adapter alone, to apply on the base model |
Usage (llama.cpp / Hermes)
# llama.cpp server
llama-server -m atanor-4b-full-Q4_K_M.gguf --jinja -ngl 99 -c 32768 --alias atanor
# point a Hermes profile at it (provider base_url: http://localhost:8081/v1)
hermes chat --profile atanor -q "read data.csv and total the 'south' region using the terminal"
Thinking is on by default — it helps the model reason about which tool to use. (Pass chat_template_kwargs: {"enable_thinking": false} to disable.)
This is my first fine-tune, and version one. More to come. I learned a ton — and the best part is it was all done at home, on my own GPU.
Built with the Hermes Agent ecosystem. Base model © Qwen, Apache 2.0.