Gemma4 MTPLX Optimized Quality

This is an MTPLX pair bundle for Gemma 4 31B speculative decoding on Apple Silicon.

It is not a single vanilla Transformers model directory. The repository contains two MLX-format artifacts:

target/ - Gemma 4 31B IT target, MLX Q8 affine group-size 64
assistant/ - official Gemma 4 31B assistant drafter, MLX Q8 affine group-size 64

Use this pair when target precision and high acceptance are the priority.

Source

Target source: google/gemma-4-31B-it
Target revision: 145dc2508c480a64b47242f160d286cff94a2343
Assistant source: google/gemma-4-31B-it-assistant
Assistant revision: cffbbd2cea41ea56a0fa5b0487e0d445121fd204

Both artifacts were converted locally to MLX format.

Quantization

Target:

bits: 8
group_size: 64
mode: affine

Assistant:

bits: 8
group_size: 64
mode: affine

MTPLX Usage

MTPLX loads the pair straight from this repository:

pip install mtplx
mtplx serve --model Youssofal/Gemma4-MTPLX-Optimized-Quality

To find the best draft block size for your Mac, run mtplx tune --model Youssofal/Gemma4-MTPLX-Optimized-Quality.

The Gemma 4 assistant is a separate drafter model. MTPLX uses exact speculative sampling with target verification and residual correction.

Local Benchmark

Prompt: single-file HTML5 Canvas Flappy Bird game, capped at 1000 generated tokens.

Sampler:

temperature: 1.0
top_p: 0.95
top_k: 64
seed: 0

Best observed block size:

block_size: 6
acceptance: 833 / 835 = 99.76%
speedup_vs_ar: 2.49x

Observed MTPLX throughput samples:

34.22 tok/s
32.88 tok/s
33.12 tok/s

The bundled benchmark JSON file is in benchmarks/.

Notes

This release is optimized for target precision and high acceptance. It is not the fastest absolute-TPS pair; for speed, use Youssofal/Gemma4-MTPLX-Optimized-Speed.

Gemma 4 is released by Google under the Gemma 4 license terms linked above.