Tag
Added native multi-token prediction (MTP) support to the exo local inference tool for Qwen3.6 MLX models, achieving up to 2x speedup on 27B models on an M5 Max laptop while maintaining exactness.