Gemma 4

The guide wires up an offline coding agent on a 64GB M1 Max: llama.cpp built with Metal, Gemma 4 in Q4, and a Multi-Token Prediction draft model that lifts generation from 58 to 72 tokens/sec.

The thread's sharpest critique wasn't the setup but the metric: every local-AI post benchmarks tokens per second, none benchmark answer quality. 'Quickly serving me slop doesn't make it more useful.'

It's a 26B Mixture-of-Experts model activating 3.8B params, fitting in 18GB VRAM when quantized — and Google admits output quality trails standard Gemma 4.

A developer who'd used a diffusion model in production said the point isn't 'smart', it's 'stupid fast' — turning prompting back into pair-programming instead of a slot machine.