Youssofal / Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed

Ternary Bonsai 2 27B MTPLX Optimized Speed

Prism ML's Ternary Bonsai 2 27B for MTPLX , with the Qwen3.8-27B draft head for speculative decoding. It is a 2-bit ternary vision-language model that runs on Apple Silicon Macs with 16 GB of memory or more. Requires MTPLX 2.12.0 or newer.

Full credit for the model goes to Prism ML . The language model and the vision tower in this pack are Prism ML's Ternary Bonsai 2 27B , carried byte for byte. Created using Bonsai by Prism ML. Bonsai 2 27B is built from Qwen3.8-27B by Alibaba Cloud. Both are licensed under Apache 2.0. Prism ML's model card and whitepaper describe the method, the benchmarks and the limits of the model itself.

What is in this pack

What MTPLX adds

Prism ML stores every projection of the language model as ternary weights in a rotated basis, so a runtime has to apply the matching Hadamard transform to the activations. MTPLX loads this model type natively, keeps the weights packed at 2 bits, and refuses a pack whose rotation metadata or vision tower is missing.

The draft head proposes tokens and the full model verifies them with exact speculative sampling, so the output follows the model's own distribution.

Speed

The draft head is on by default (MTP speculative decoding).

Measured on an M5 Max with 128 GB on 22 September 2026 at temperature 1.0, top-p 0.95 and top-k 20, thinking off, 512 generated tokens: depth 1 decodes 50.0 tok/s on a 4K prompt and 42.6 on a 16K prompt, against 38.8 and 34.3 with plain decoding. Depths 2 and 3 were slower. The 4-bit Qwen 3.8 27B measured 49.8 to 52.7 and 49.4 to 50.1 tok/s in the same session, so Bonsai matches it at 4K and is slower at 16K, in half the memory.

How to run it

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

The served id is mtplx-bonsai-2-27b-optimized-speed . In the Mac app, pick Bonsai 2 27B Optimized Speed. The app and the CLI recommend it first on M3, M4 and M5 Macs with 16 to 31 GB of memory.

Recommended settings

These are Prism ML's and Qwen's recommendations. MTPLX uses the thinking settings by default.

Reasoning effort is medium by default. xhigh is also available and thinks for much longer. Prism ML states that low is not supported.

Memory

On a 16 GB Mac the window is too small for the system prompt of an agent client such as OpenCode (about 18,700 tokens); use 18 GB or more for those.

License and attribution

Apache 2.0. LICENSE and NOTICE.txt are carried verbatim from Prism ML's pack. This pack contains Prism ML's Ternary Bonsai 2 27B (Copyright 2026 Prism ML, Inc.), which is built from Qwen3.8-27B (Copyright 2026 Alibaba Cloud), and the Qwen3.8-27B draft head as packed for MTPLX.