Gemma 4 is now up to 90% faster on Apple Silicon in Ollama 0.31

Ollama 0.31 introduces a substantial speed boost for Gemma 4, with Apple Silicon users seeing nearly 90% faster token generation in coding-agent benchmarks.

The performance gain comes from a multi-token prediction method that proposes and verifies several tokens at once, with the draft length auto-tuned in real time and no change to model outputs.

An optimized matrix multiplication kernel contributed to MLX further accelerates batch verification, benefiting both Gemma 4 and other models.