Gemma 4 31B MTP Runs on MacBooks

The interesting part here is not the absolute number, but the gap between Google’s promised MTP speedups and this first real-world local result.

• –10-12 tok/s on an M5 MacBook with 128GB is respectable for a 63GB BF16 dense model, but not an obvious breakthrough.
• –The test is being run through Ollama/Open WebUI, so the number is framework-dependent and may not reflect the best possible MTP-aware performance.
• –llama.cpp support is still missing here, which limits how broadly this can be evaluated right now.
• –The real question is comparison against existing local coding defaults on the same hardware, especially Qwen3.6 27B and Qwen3 Coder Next.
• –The post reads as an early signal, not a verdict: promising model, but not yet enough evidence to justify a switch.