research

• 2026-07-21 · engineering · research · 8 min

  Chunking Qwen3.5's gated DeltaNet for 1.5x faster prefill on Apple Silicon

  Qwen3.5's hybrid layers replace attention with a gated delta rule, a sequential state update torch.compile cannot run fast. How Alloy maps it onto Metal: one serial kernel for decode, and a two-stage chunked kernel that makes prefill 1.5x faster at production depth.

• 2026-07-14 · engineering · research · 5 min

  How to run YOLO11 on the Coral Edge TPU with 3x less quantization loss

  Ultralytics' Edge TPU export quantizes the whole detection graph and quietly pays for it in accuracy. Cutting the model at the raw heads keeps 3x more of it.