Qwen leads local model rankings for Mac

Alibaba’s Qwen 3.6-35B-A3B leads LLMCheck’s ranking of 116 open-weight models for Apple Silicon as of 24 October 2026, scoring 81. Recursive Flint follows with 80, ahead of KAT-Coder-V2.5 at 79. The leading model requires 24 GB of unified memory and has a published decoding speed of 118 tok/s on an M5 Max.

The composite ranking combines benchmark capability, decoding speed, memory accessibility and licence openness. Speed figures carry provenance labels identifying published results, community measurements or estimates. Reference performance generally reflects single-stream decoding on an M5 Max, with an M5 Ultra used for models requiring more than 128 GB. Measured runs replace estimates where available, and the score uses the displayed speed.

Measurement coverage remains limited: 193 measured rows cover 48 models, but only 26 of the 116 models use measured results for their leaderboard speed. Maple Preview 20B-A1B posts the fastest measured figure at 387 tok/s on an M5 Pro using ternary weights. Hardware-specific recommendations also account for whether a model fits available memory, meets the speed threshold and has a standard Mac runtime.