GreenBit adds a speculative decoding head to ternary Bonsai 2 27B

GreenBit AI has released a serving bundle for Prism ML's Ternary Bonsai 2 27B that combines unusually low-bit model weights with a model-specific multi-token prediction head. The new Hugging Face repository distinguishes two components: Prism's original ternary model assets, which GreenBit says are unchanged, and an optional draft head stored under mtp/. The repository declares Apache-2.0 and documents the bundled files and notices. This is principally an inference engineering release, not evidence that a new 27B base model was trained.

Ternary weights represent values using minus one, zero and plus one, enabling compact packing. The additional head proposes several future tokens; the target model then verifies those proposals before accepting output. That matters because speculative decoding can reduce repeated full-model passes without simply trusting a smaller model's answers. GreenBit describes distilling the head on Bonsai-generated prose, code, chats and coding-agent examples, beginning from a Qwen3.8-27B multi-token prediction component.

In its own matched Mac Studio M3 Ultra test, GreenBit reports 114 decode tokens per second on short prompts with its gbx-lm 0.8.2 implementation, against 42 for Prism's referenced MLX runtime. On a DGX Spark it reports 62 tokens per second and 5.9 GB of loaded model memory, compared against the configurations listed in its repository. These are publisher-conducted measurements, not independent Open Model Weights results. Prompt length, decoding mode, exact quantization, kernel versions and cache policy can change the comparison.

The practical takeaway is a concrete, inspectable approach to faster local 27B inference with a distinct optional acceleration component. Teams can test the OpenAI-compatible local endpoint, compare draft-head-on versus off using identical prompts, and measure end-to-end throughput, output equality, memory and latency on their own hardware. Do not generalize the reported ratios to all GPUs or assume that a compact weight file represents peak application memory.