Commit History

  1. Add MoE DFlash2 A/B: loses harder than dense (0.89x code), contradicting the prediction

    Verified signature
    458cf03
  2. Add DFlash2 speculative decoding A/B on dense 27B (net loss despite 76% acceptance)

    Verified signature
    f2573a0

  1. Add oMLX Qwen3.8-27B dense 4bit at 8k ctx (2 sessions, GPU idle both times)

    Verified signature
    ea246f1

  1. Add M2 Max 32GB benchmarks: MoE vs dense, MTP on Metal, KV quant, MLX vs llama.cpp

    Verified signature
    9b393f1