MLX LoRA adapter checkpoints (training steps 100–500) fine-tuning Gemma-4-12B for Eyla agent tool-calling, plus a merged Q4_K_M GGUF (eyla-gemma4-q4km.gguf).

Honest status: in our agent evals the stock Gemma-4-12B QAT outperformed this fine-tune (7/8 vs 1/8 tasks). Kept public for reproducibility; if you want a working Eyla tool-caller, use eyla-qwen3-8b-tools-v2 instead. Gemma license terms apply.