Until now, anyone who had fine-tuned a language model in BF16 had no public way to convert it into Prism ML's ternary format. OpenBonsai, an Apache-2.0 open-source framework, handles that conversion and produces files compatible with the PTQ1_0 GGUF format. The project grew out of a practical need: its author works on WebBrain, an open-source browser agent, and aims to run as much intelligence as possible locally, ideally through WebGPU.
The starting point is Prism ML's Ternary Bonsai, which compresses a 9-billion-parameter model to about 2 GB using ternary weights of -1, 0, and +1. Prism released the weights and the inference runtime but not the full BF16-to-ternary compression pipeline. Users could fine-tune the existing ternary models, but there was no public method to bring their own full-precision fine-tune into that format. OpenBonsai fills that gap.
According to the project documentation on GitHub, the framework compresses a dense model into ternary weights in the PTQ1_0 GGUF format, which runs in Prism's llama.cpp fork. For Qwen3.5-9B, the output is a file of about 2 GB, with the same tensor layout as Ternary-Bonsai-2-9B. The quantization is post-training, so the model does not have to be retrained from scratch.
The context explains why this matters. AlexTech had already covered Ternary Bonsai 2 27B, which keeps 98.2% of the benchmark performance of its Qwen3.8 27B base in a 5.9 GB footprint. Those models could be used, but not recreated from a user's own version. With OpenBonsai, compression becomes a step anyone can repeat on their own fine-tune, although the project is still experimental, as its repository states.
The main beneficiaries are developers and researchers working on custom models who want to deploy them on low-memory devices, from laptops to browsers. The framework is not a push-button solution: the quality of the result depends on the starting model and the procedure applied, and the documentation does not report systematic comparisons between the original model and the compressed version.
The next step indicated by the project is extending the framework beyond Qwen3.5 to other architectures, while consolidating the pipeline toward Prism's format. For anyone following the race for lightweight local models, OpenBonsai turns ternary compression into an open process rather than something available only in the released weights.
No comments yet. Be the first!