PrismML has released Bonsai 2 27B, a compact language model based on the open-source Alibaba Qwen3.8 27B. The new version occupies 5.9 GB of storage, which is 9 to 10 times smaller than the original model, allowing it to run on standard PCs and flagship smartphones.
The startup utilizes a proprietary compression technology that reduces the size of model weights. By shifting from standard 16-bit weights to ternary weights—represented as +1, 0, and -1—the company significantly decreases the memory footprint required for the system.
According to the developer, Bonsai 2 27B achieves 98% of the benchmark performance of the original Qwen model. This is an improvement over the company's previous iteration released in March, which reached 95% of the original performance.
PrismML reports that the original Qwen model has been downloaded 11 million times, while the company's compressed versions have reached 2.6 million downloads. The team plans to apply this compression method to larger models with hundreds of billions of parameters.
