MTPLX: the fastest way to run Qwen 3.8 on a Mac. Native multi-token-prediction speculative decoding on Apple Silicon, two to three times the speed of plain decoding, exact at any temperature.
Qwen 3.8 27B Bare Speed FP16
Quickest burst chat speeds. Lower quality and slower on long coding tasks. This is the M1 and M2 build.
The FP16 precision sibling of Qwen 3.8 27B Bare Speed. M1 and M2 Macs do not run bf16 well, so this build keeps every quantized weight byte-identical to the parent and stores the remaining floating tensors (scales, biases, norms, the GDN convolution and state parameters, and the MTP head) in fp16 instead of bf16. Same layout, same tuned depth and draft settings, same context window. On an M3 or newer Mac use the parent instead.
MTPLX picks the right one for you: the app and
mtplx start route M1 and M2 Macs to the FP16 builds and everything newer to
the parents. Draft sampler temperature 0.6 is stamped in, the measured winner for this build.
Measured on MTPLX 2.11.3 (16 September 2026)
This is a Qwen3.8-27B MLX pack for MTPLX, the fastest way to run Qwen 3.8 27B on a Mac. All numbers on a MacBook Pro M5 Max, fans verified at maximum, sampled at the model's own settings. Conditions and sources: mtplx.com/benchmarks.
| Run | tok/s |
|---|---|
| Optimized Speed rewriting a file it just wrote, stock settings (MTPLX 2.10.0) | 87.6 |
| Bare Speed on a fresh coding task at official Qwen 3.8 sampling (MTPLX 2.7.0) | 65.2 |
| Optimized Speed on a 3k-token chat answer (MTPLX 2.10.0) | 64.3 |
| The 27B record on a fresh generation: Qwen 3.6 27B Optimized Speed, 192-token bench, 2 July 2026, raw logs published | 81.74 |
Pack quality against the bf16 Qwen3.8-27B checkpoint, teacher-forced over 2,389 positions of code, prose, JSON and a multilingual notice (16 September 2026): Optimized Speed (4-bit dynamic) 96.0 percent top-1 agreement and KL 0.012; Optimized Quality (8-bit dynamic) 99.3 percent and KL 0.0005. Exactness on this release: a thousand four-token draws from the fast path match a thousand from the plain path within the plain path's own noise, at temperature 1, top-p 0.95, top-k 20. Details: MTPLX 2.11.3 release notes.
Runs on Apple Silicon Macs with 32 GB of unified memory or more (36 GB for Optimized Quality): MacBook Pro, MacBook Air, Mac mini and Mac Studio on M1 to M5. Guide: Run Qwen 3.8 27B on a Mac. Comparison pages: MTPLX vs mlx-serve, MTPLX vs oMLX, MTPLX vs LM Studio, MTPLX vs Ollama.
Speeds
The numbers we publish for the parent were measured on an M5 Max: 65.2 tok/s on the coding task and 32.4 tok/s sustained over a single 52,740-token answer, official Qwen 3.8 sampling, generation running to the model's own stop. This FP16 build has the same weights and runs the same MTPLX turbo path, so the speculative math is identical; absolute tok/s on an M1 or M2 depends on that chip. We have not published M1 or M2 numbers for it yet.
| Download | 16.0 GB |
| Peak unified memory (parent, measured on M5 Max) | 17.0 GB |
| Context window | 262,144 tokens |
| MTP depth | 3 |
| Sampling | temperature 1.0, top-p 0.95, top-k 20 (the official Qwen 3.8 contract) |
MTPLX_FP16_CONVERSION_MANIFEST.json in the repo lists every tensor that was
cast and every tensor that was preserved, with sha256 for each shard.
Speculation in MTPLX is exact at any temperature: drafts are accepted with the
probability-ratio rule plus residual resampling.
Use it
Mac app: download at mtplx.com, pick "Qwen 3.8 27B Bare Speed FP16".
Command line:
pip install mtplx
mtplx serve --model Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed-FP16
The other two FP16 builds: Optimized Speed FP16, Optimized Quality FP16.