MTPLX: the fastest way to run Ternary Bonsai 2 27B on a Mac. Native multi-token-prediction speculative decoding on Apple Silicon, exact at any temperature.

Bonsai 2 27B Optimized Speed

Prism ML ternary 27B model with vision and MTP. Compact weights for smaller Macs.

Prism ML's Ternary Bonsai 2 27B is Qwen 3.8 27B rebuilt with ternary weights, and Prism ML reports it keeps 98.2 percent of the full model's benchmark average in 8.85 GB. On MTPLX it runs on GPU kernels written for it, with the Qwen 3.8 27B multi-token-prediction head drafting ahead. That makes it faster than the 4-bit Qwen 3.8 27B in half the memory. It reads images, and it runs on a Mac with 16 GB.

Prism ML's own figure for Bonsai 2 27B on an M5 Max with MLX is 46.8 tok/s. On MTPLX the same chip decodes it at 64.4 tok/s after a 4,061-token prompt.

Measured on MTPLX 2.12.0 (22 September 2026)

This is the Bonsai 2 27B MLX pack for MTPLX, the fastest way to run Bonsai 2 on a Mac. MacBook Pro M5 Max with 128 GB, sampled at the model's own settings (temperature 1.0, top-p 0.95, top-k 20), thinking off, 512 generated tokens, one model loaded at a time.

Prompt	Bonsai 2 27B Optimized Speed	Qwen 3.8 27B Optimized Speed	Peak memory, Bonsai and 27B
4,061 tokens	64.4 tok/s	52.6 tok/s	11.4 and 23.9 GB
16,350 tokens	57.1 tok/s	51.0 tok/s	14.6 and 27.0 GB

Runs on Apple Silicon Macs with 16 GB of unified memory or more. From MTPLX 2.12.0 it is the first recommendation on M3, M4 and M5 Macs with 16 to 31 GB.

How it is built

Prism ML stores every weight of the language model as −1, 0 or +1 with one scale per group, so the 27B language model takes 7.67 GB.
MTPLX runs those weights natively, with GPU kernels for Prism ML's rotation and for the ternary matrix multiply.
The Qwen 3.8 27B multi-token-prediction head drafts ahead, and the full model verifies its drafts in one pass.
The Qwen 3.8 27B vision tower is included.


Download	8.85 GB
Peak while serving	11.4 GB at a 4K prompt, 14.6 GB at 16K
Recommended Macs	16 GB+ unified memory (18 GB+ for agent clients such as OpenCode)
Context window	8,192 tokens on 16 GB, 20,480 on 18 GB, 94,208 on 24 GB, up to 262,144
MTP depth	1
Reasoning effort	medium by default, xhigh available
Sampling	temperature 1.0, top-p 0.95, top-k 20

The serving contract ships inside mtplx_runtime.json. MTPLX reads it on load. Drafts are accepted with the probability-ratio rule plus residual resampling, so the output follows the model's own distribution at any temperature.

Use it

Mac app (MTPLX 2.12.0 or later): download at mtplx.com, pick "Bonsai 2 27B Optimized Speed".

Command line (MTPLX 2.12.0 or later):

pip install mtplx
mtplx serve --model Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed

Siblings: Qwen 3.8 27B Optimized Speed (the 4-bit 27B, for Macs with 32 GB or more) and MiMo V2.6 Qwen 9B Optimized Speed.

Base model: prism-ml/Ternary-Bonsai-2-27B-mlx-2bit, built from Qwen/Qwen3.8-27B (Apache 2.0; LICENSE and NOTICE.txt in this repo). Created using Bonsai by Prism ML.