I re-trained the DFlash 2 drafter for Ternary Bonsai 2 27B: 2.2x on an L4 (3.2x on code edits with ngram lookup), 1.5x on a Mac, 1.2x in Chrome

PrismML's Ternary Bonsai 2 27B fits a 24 GB card or Mac, but decodes at ~30 tok/s on an L4 and ~21 on an M4 Pro. z-lab's DFlash 2 drafter was trained on bf16 Qwen3.8-27B, so it guesses worse on the ternary model. I fine-tuned it on 1.5M tokens of Bonsai 2's own greedy output.

• Drafter (safetensors + Q4_K_M GGUF): https://huggingface.co/naklitechie/Qwen3.8-27B-DFlash2-ternary-bonsai2
• Setup for all three runtimes: https://github.com/NakliTechie/dflash-mlx-bonsai2/blob/main/docs/USAGE.md

NVIDIA (PrismML's llama.cpp fork, prism branch):

llama-server -m Ternary-Bonsai-2-27B-PQ2_0.gguf -md Qwen3.8-27B-DFlash2-r3-Q4_K_M.gguf \ --spec-type draft-dflash --spec-draft-n-max 7 --spec-type ngram-mod \ -ngl 999 -ngld 999 -fa on --jinja 

One L4, greedy: GSM8K 2.17x, MBPP 2.17x, MATH-500 2.20x, MT-Bench 1.39x. Code edits: 3.15x with ngram-mod stacked (drafter alone 2.46x). Accuracy within 1-2 problems per set.

Mac: a fork of bstnxbt/dflash-mlx with an 8-row 2-bit Metal GEMM for the verify step. M4 Pro: 1.5x on raw code completion, 1.3x on chat code, 1.2x on math. One script gives you an OpenAI-compatible server.

Browser: a WGSL port inside LocalMind (https://localmind.naklitechie.com), on by default for Bonsai 2 27B. 1.18x on code, output identical.

Chat and prose are about break-even. Use temperature 0.

Credit to z-lab (DFlash 2), PrismML (Bonsai 2, llama.cpp fork) and bstnxbt (dflash-mlx). Numbers are from one L4 and one M4 Pro; results from a 3090, 4090 or other Apple chips are welcome.