MTPLX 2.5.1 makes Qwen 3.6 27B Optimized Speed V2 the recommended coding model on modern Macs with enough memory.
Optimized Speed V2 is much higher quality than the original Optimized Speed model. It uses dynamic 4-bit hybrid quantization with hand-tuned sensitive parts kept at up to 16-bit. In real coding QA, it performed better as agent work became longer.
The tradeoff is straightforward. V2 is slightly larger and can be a little slower for short chat turns. The original Optimized Speed model remains fully supported and appears directly below V2 for users who prefer the smaller download or mostly use short chats.
V2 is now wired through the complete product path:
The app and CLI use the same memory-aware policy. Modern Macs with at least 32 GiB of detected unified memory get V2 first. Smaller Macs keep the existing 9B and 4B recommendations.
Open issues and pull requests were reviewed before the release. None had a better benefit-to-risk ratio than this focused model launch. The open feature pull requests are broad or experimental, while the active cache and cross-hardware performance issues need their own measured work. No unrelated architecture change was pulled into 2.5.1.
Runtime kernels, sampler defaults, and speculative depth are unchanged from 2.5.0. The release also fixes a cold-cache bookkeeping race so flushes wait for in-flight writes and eviction decisions use a coherent disk snapshot.
pip install -U mtplxbrew upgrade mtplx