The pack
Optimized Speed is the
speed-focused 35B checkpoint. MTPLX uses the model's own MTP heads to generate draft tokens, then verifies them with
the main model; when the draft heads are well matched you get higher throughput without running a separate drafter
model. This checkpoint is optimized for that path: the MTP heads are calibrated INT4, shipped as a separate
mtp.safetensors sidecar next to the 4-bit body shards. MTPLX reads mtplx_runtime.json and
selects the measured defaults automatically.
A sibling pack, Optimized
Balance, and FP16 siblings of both for M1 and M2 Macs (same weights, native precision for chips without bf16)
live in the same namespace. The served model id for this pack is mtplx-qwen36-35b-a3b-optimized-speed.
Speed on your Mac
MTPLX publishes no headline tok/s for this pack 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/Qwen3.6-35B-A3B-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.
Three dated facts about this pack's speed path. MTPLX 2.2.0 (18 July 2026) moved its default draft depth from 3 to 2 after measuring both on an M5 Max. MTPLX 2.4.0 (31 July 2026) compiled its decode stack. MTPLX 2.6.0 (11 August 2026) added concurrent speculative decoding: two agents at once decode at 1.6 to 2.25x per lane against the previous AR batch route on the 35B-A3B, M5 Max, sampled at shipped settings.
RAM and Macs
- Memory tier: MTPLX publishes no RAM tier for this pack here. 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).
- M1 and M2: the FP16 siblings carry the same weights at native precision for chips without bf16.
- Image input: PNG, JPEG and WebP with MTP intact since 1.0.3 (11 June 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 Qwen 3.6 35B-A3B Optimized Speed pack. The app downloads it, 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/Qwen3.6-35B-A3B-MTPLX-Optimized-Speedmtplx run "hello" --model Youssofal/Qwen3.6-35B-A3B-MTPLX-Optimized-Speed answers one prompt in the
terminal. Then 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. Since 2.6.0 one server takes concurrent requests with speculation on, so two agents can share it.
How it is built
| Component | Format |
|---|---|
| Main body | 4-bit MLX affine, 64-weight groups |
| Router and gate tensors | 8-bit where recorded by the config |
| MTP numbered-expert weights | Calibrated INT4, 32-weight groups, in mtp.safetensors |
| Norms, scales, biases, plain tensors | bf16 |
The runtime contract in mtplx_runtime.json records the backend (qwen3-next-mtp), the
sampler for target and draft (temperature 0.6, top-p 0.95, top-k 20), and the Sustained profile. Activations stay bf16
on every pass (fp16 on M1 and M2); the speculative path runs the same model as the autoregressive path. This is a
quantized checkpoint built for fast local use on Apple Silicon through MTPLX.
Exactness
Speculation in MTPLX is exact. Drafts from the MTP heads are accepted with the probability-ratio rule and rejected drafts are resampled from the residual, so what you sample is what the model would have sampled without speculation, at any temperature. The draft sampler is a speed knob only. The quantization is the one approximation.