This 3.14 GB GGUF claims 43% less math overthinking—without another training run
A community weight edit of Prism ML’s Bonsai targets repeated second-guessing. The author reports shorter reasoning on sampled tests, with slower per-token decoding and an instruction-following tradeoff.
A new community GGUF edit aims to stop a small reasoning model from repeatedly changing its mind. Hugging Face author luvette says Ternary-Bonsai-2-9B Decisive V2 uses 43% fewer thinking tokens on a 250-question GSM8K sample and 27% fewer on a 250-question MATH-500 sample, without a new training run.
The 3.14 GB main file modifies Prism ML’s existing Bonsai model, itself derived from Qwen3.5-9B. The author describes finding directions associated with reopening decisions and re-checking math, then projecting them out of selected weights. The edited matrices use higher-precision storage and projection-aware rounding; this is a weight edit, not a new fine-tuning run or an external instruction telling the model to be concise.
Shorter thinking is the hook, not a universal accuracy win
The card reports MATH-500 accuracy rising from 75.2% to 81.2% under a 12,288-token budget. Those figures cover the sampled 250 problems, not the full benchmark. The author attributes the gain mainly to finishing more problems before running out of tokens. On the 200 problems both models finish, reported accuracy is 93.5% for the base and 93.0% for V2.
There are costs. Sampled GSM8K accuracy is 94.4%, versus 95.2% for the base. Strict prompt-level IFEval falls from 92.7% to 90.3%; the author flags possible trouble with exact length constraints, while noting uncertainty in that difference. Per-token decoding is also about 12% slower than the base in the reported single-stream test, and the file is larger. AITrending has not independently reproduced these results.
A local route, with two edits to choose from
The main q6pa-m version removes both habits. A same-size q6pa alternative removes only decision reopening and reports better preservation of instruction following, at the cost of longer thinking. Optional 3.39 GB MTP files add a draft head for compatible llama.cpp speculative decoding.
The English guide documents stock llama.cpp and Ollama 0.40.14, with tested Linux RTX 3070 and macOS Apple M4 configurations for Ollama. Older Ollama engines need support for Bonsai’s Q2_0 format and Qwen3.5 architecture; Ollama does not use the MTP head. Downloading the ordinary GGUF and importing the supplied Modelfile provides the documented Ollama route.
The release is Apache-2.0 with Bonsai and Qwen attribution. Vision after the edit remains unevaluated, and the author has not yet compared it with cheap alternatives such as a concise prompt or a hard thinking budget. For local-model experimenters, the useful question is whether fewer repeated reasoning tokens outweigh those tradeoffs on their own tasks.
