PrismML has released Ternary Bonsai 2 27B, a compressed version of Qwen3.8 27B that cuts the checkpoint from around 54GB down to 5.9GB. The weights are published on Hugging Face under the Apache 2.0 licence, so the model is free to use, modify and deploy commercially without the restrictions that come with research-only licences.
This is not a new model trained from the ground up. PrismML takes the existing Qwen3.8 27B weights and rewrites each one as a ternary value, meaning every weight becomes -1, 0 or +1, at roughly 1.7 bits per weight rather than the 16 or 32 bits typical of an uncompressed model. That is what drives the size drop: a nine-times smaller file for the same underlying model.
The trade-off question for any compression technique is how much capability survives it. PrismML reports that Ternary Bonsai 2 retains 98.2 per cent of Qwen3.8 27B's benchmark average, which, if it holds up under independent testing, is a strong result for this level of compression. The full checkpoint carries 27.36 billion parameters and a 262,144-token context window, with a separate 0.63GB vision component loaded only when images are part of the input.
It follows the first Bonsai 27B, released in July 2026, which kept 95 per cent of its base model's performance at the same ternary compression level. The jump to 98.2 per cent in this second generation suggests PrismML's quantisation method is maturing. For anyone running large models on constrained hardware, a 9x size reduction with minimal quality loss is a meaningful option, though the benchmark figures are vendor-reported and worth checking independently before deployment decisions.
Source: MarkTechPost
