0.16.7 — Faster DeepSeek V4.1 Flash decoding

Up to ~1.9× faster DeepSeek V4.1 Flash decoding on the GPTQ v2 checkpoint in a same-machine M3 Ultra 256 GB, single-stream comparison: 17.6–33.9 tok/s on the 0.16.6 PyPI build rose to 32.6–44.9 tok/s on the 0.16.7 release candidate across five prompts. The published 0.16.7 wheel was not measured separately. Reproduction and measurements.

• DeepSeek V4.1 Flash. Fused draft kernels, shared vocabulary-head reads during multi-row verification, less draft overhead, and mixed-bit fused MoE speed up the experimental text-only, greedy-only build. The default draft depth is 2; depth 4 or auto can be chosen for code-heavy workloads. The 224 GB catalog floor is enforced, and the default weights require a 256 GB Mac. (#5072, #5041, #5059, #5048)
• Desktop updates. Update availability refreshes while the app is open, on return when stale, after failures with bounded retries, and on an explicit check. The offer uses a compact strip. (#5062)
• Dependency repair. Optional runtime, connector, and audio install hints target the Python interpreter running Rapid-MLX. (#5060, #5064)