Dispatch
Ollama v0.31.2 Speeds Gemma 4 ~90% on Apple Silicon via Auto-Tuned Multi-Token Prediction
Ollama v0.31.2 (July 6, 2026) delivers nearly 90% faster Gemma 4 token generation on Apple Silicon through multi-token prediction with automatic draft-length tuning — on by default, no config, and output-preserving. It also adds MLX-engine support for Command A and North-family models on Macs, flash attention for older NVIDIA GPUs (compute capability 6.x), and iGPU vision-model offload. For builders running local models on Macs the speedup is a free throughput win, and Ollama remains the #1 trending AI repo (~176K stars).
Source
↳ Follow the thread