35 domain-expert LoRAs on a MoE (Qwen3.6-35B-A3B) trained with MLX on one Mac Studio — what worked, what didn't, all Apache-2.0
Over the last months I trained 35 domain-expert LoRA adapters on Qwen3.6-35B-A3B (native MoE, 256 experts, 3B active) with MLX on a Mac Studio M3 Ultra 512 GB, BF16, no quantization during training. Everything is public: adapters, datasets, per-domain configs, router, training scripts.