Two packs
The Optimized Speed repository is a pair bundle holding two MLX artifacts: a target/ folder with the
Gemma 4 31B IT model and an assistant/ folder with Google's official Gemma 4 31B assistant drafter, both
converted locally to MLX format. Its card points to Optimized Quality for a higher-precision target.
| Pack | Target | Drafter | Pick it for |
|---|---|---|---|
| Optimized Speed | Gemma 4 31B IT, MLX 4-bit affine, 64-weight groups | Gemma 4 31B assistant, MLX 6-bit affine, 64-weight groups | Absolute throughput. |
| Optimized Quality | Higher-precision target | A higher-precision target. |
The served model id for Optimized Speed is mtplx-gemma4-optimized-speed. Gemma 4 is released by
Google under the Gemma 4 license terms.
Speed on your Mac
MTPLX publishes no tok/s for this pair with full conditions. The speed is measured on your
own Mac during onboarding: auto-tune runs the model itself at each draft depth with fans pinned, keeps plain autoregressive
decoding as the baseline, and saves a depth only if it beats it. Re-run it any time with
mtplx tune --model Youssofal/Gemma4-MTPLX-Optimized-Speed --retune. The live dashboard
(mtplx start dashboard) shows decode speed and acceptance by depth while you work, and
mtplx start --no-mtp runs plain autoregressive decoding on the same loaded model for a
side-by-side check.
RAM and Macs
- Memory tier: MTPLX publishes no RAM tier for this pair here. Both the 31B target and the 31B drafter load together. The app checks your Mac before recommending anything, and the memory governor (2.10.0) prints engine budget, weights, resolved context window and session bank in the serve banner. Requests that cannot fit are refused up front with HTTP 507 (2.10.2, 1 September 2026).
- Other tiers: Qwen 3.8 27B is the default on 32 GB or more; Qwen 3.5 9B and 4B cover 16 GB.
Install
Mac app: download the DMG and pick the Gemma 4 Optimized Speed pack. The app downloads both artifacts, sets up its engine, and measures your machine to pick the fastest decoding depth.
Command line:
brew install youssofal/mtplx/mtplx
mtplx serve --model Youssofal/Gemma4-MTPLX-Optimized-SpeedThen point OpenCode, Pi, Claude Code, Cline, Cursor or anything that speaks the OpenAI or Anthropic API at
http://127.0.0.1:8000; mtplx connect claude-code and mtplx connect opencode print
the exact client config, and setup pages for each client are in the docs. The Sustained profile is
the default for this pack.
To benchmark the pair directly after downloading the repository, the card's command points MTPLX at the two subdirectories:
mtplx bench gemma-mtp \
--target-model ./target \
--assistant-model ./assistant \
--prompt-suite mtplx/benchmarks/prompts/flappy.jsonl \
--max-tokens 1000 \
--draft-block-sizes 6 \
--allow-unverified-gemmaHow it is built
| Artifact | Source | Format |
|---|---|---|
target/ | google/gemma-4-31B-it, revision 145dc2508c480a64b47242f160d286cff94a2343 | MLX affine, 4-bit, 64-weight groups |
assistant/ | google/gemma-4-31B-it-assistant, revision cffbbd2cea41ea56a0fa5b0487e0d445121fd204 | MLX affine, 6-bit, 64-weight groups |
Both artifacts were converted locally to MLX format. Unlike the Qwen packs, the draft head is a whole separate model: Google's own assistant drafter for Gemma 4 31B. Activations stay bf16 on every pass (fp16 on M1 and M2), and the target that verifies is the same model that would answer without speculation.
Exactness
Speculation in MTPLX is exact. Drafts from the assistant are accepted with the probability-ratio rule and rejected drafts are resampled from the residual, so what you sample is what the target would have sampled without speculation, at any temperature. The quantization of the target is the one approximation.