What changed
Signed prereleases b11059-b11064 add F16 Metal FWHT support while unifying capability and dispatch checks that could otherwise abort, make a Mamba time-step projection contiguous, enable sparse CUDA flash attention for Qwen4, handle invalid UTF-8 in the PEG syntax tree, and keep more Kimi-K3 cross-layer residual work on Metal instead of falling back to CPU.
Why it matters
The sequence combines a concrete Metal abort prevention and parser hardening with new model and accelerator paths. These changes can affect correctness, backend selection, latency and memory on exact workloads rather than only headline benchmark speed.
What to do
Canary b11064 beside the last approved stable build on the exact model, quantization, GPU and driver you operate. Exercise an affected F16 Hadamard path, Mamba projection, Qwen4 sparse attention, malformed UTF-8 grammar input and Kimi-K3 checkpoint depths; compare outputs, backend placement, memory and latency, then prove rollback.
Caveat
Every cited build remains a fast-moving prerelease. The project supplies targeted tests and one M5 Pro microbenchmark, but broad cross-model and cross-device correctness and performance remain independently unverified.