The fastest way to run Qwen 3.8 on a Mac.

MTPLX is a native Mac app and a command line that runs local language models on Apple Silicon with the model's own multi-token prediction (MTP) heads. It runs Qwen 3.8 Flash Next, the 125B mixture of experts, Qwen 3.8 27B and Prism ML's Ternary Bonsai 2 27B, plus Qwen 3.6, Qwen 3.5 and Gemma 4. The model drafts several tokens ahead of itself, one batched forward pass verifies the draft, and tokens are committed through exact rejection sampling with residual correction. The sampler works at the model's native temperature, and decode runs at around twice the speed of plain decoding: measured 1.6x on a 16 GB M4 Mac mini and 2.24x on an M5 Max.

Measured speeds

Every number below was measured on a MacBook Pro M5 Max with 128 GB, fans verified at maximum, sampled at the model's own settings. Conditions and sources for every row, with the raw logs where they exist, are at mtplx.com/benchmarks.

Model	Speed	Run
Qwen 3.8 Flash Next, Optimized Speed	63.5 tok/s	65,502-token prompt read at 1,094 tok/s (768 on 2.11.3), first token after 60.1 s instead of 85.5 s, 512 tokens generated, thinking off, MTPLX 2.12.0 (56.1 tok/s on 2.11.3)
Ternary Bonsai 2 27B, Optimized Speed	64.4 tok/s	4,061-token prompt, 512 tokens generated, MTP depth 1, thinking off, MTPLX 2.12.0 (52.6 for the 4-bit Qwen 3.8 27B in the same session), 11.4 GB peak
Qwen 3.8 Flash Next, Optimized Speed	125.8 tok/s	one OpenCode request: 1,301 tokens generated, 18,539-token prompt with 18,364 tokens served from cache, MTP depth 3, MTPLX 2.11.3, 16 September 2026
Qwen 3.8 Flash Next	79.3 tok/s	9k-token code prompt, 1,500 tokens generated, thinking off, two alternating boots each, MTPLX 2.11.3 (62.5 on 2.11.2)
Qwen 3.8 Flash Next	61.8 tok/s	109k-token OpenCode turn, mean of two runs, MTPLX 2.11.3
Qwen 3.8 Flash Next	50.3 tok/s	200k-token OpenCode turn, warm, mean of two runs, MTPLX 2.11.3
Qwen 3.8 27B, Optimized Speed	87.6 tok/s	rewriting a file it just wrote, stock settings, MTPLX 2.10.0
Qwen 3.8 27B, Bare Speed	65.2 tok/s	fresh coding task at official Qwen 3.8 sampling, generation to the model's own stop, MTPLX 2.7.0
Qwen 3.6 27B, Optimized Speed	81.74 tok/s	the 27B record on a fresh generation: 192-token bench, thinking off, temperature 0.6, twin runs, 2.69x over 30.37 plain decode, 2 July 2026, raw logs
Qwen 3.5 4B, Optimized Speed	227.8 tok/s	depth 3, 1.71x over 133.6 plain decode, MTPLX 2.2.0

MTPLX uses rejection sampling with residual correction at the model's native temperature. For 2.11.3, MTPLX compared a thousand four-token samples from the fast path with a thousand from the plain path at temperature 1, top-p 0.95 and top-k 20, by token id. Those measurements matched within the plain path's own noise on both Flash Next and the 27B Quality pack. From 2.12.0, image requests also use the compiled verifier, and that sampled comparison has not yet been repeated for images. The acceptance rule follows the Leviathan and Chen rejection sampling theorem; numerical equivalence of the model's execution paths is checked separately. There is no second draft model eating your RAM, and no greedy shortcut that quietly changes what the model would have said.