Fused MLX 4-bit model for local-first tool calling in Clawdia-style chat.
It was trained from Qwen3 8B with LoRA/QLoRA data focused on:
This is the direct MLX inference artifact. You do not need to load a separate
adapter.
1python -m mlx_lm chat \
2 --model spanthee/qwen3-macos-clawdia-toolcalling-mlx-4bit \
3 --temp 0.1 \
4 --top-p 0.9 \
5 --max-tokens 1200
For tool-calling use, include the available tool list and schemas in the prompt.
The model was trained to treat the tool list in the current prompt as
authoritative rather than relying on memorized tool names.
Current local evidence reports 184/184 included tool-calling eval passes with
strict JSON output on the included rows. Some reported passes use harness-level
postprocessing for runtime-safe normalization, so raw-output behavior should be
validated in the target Clawdia runtime before treating the model as complete.
Known weak area: multimodal receipt-image extraction is not covered by this
text-only Qwen artifact. Receipt-image support is being evaluated separately
with Gemma VLM.
