According to CoinMeta, PrismML released a model that compresses 27 billion parameters (the size of Qwen3.8) to just 5.9GB. The weights of the model are compressed into three values: -1, 0, and +1, resulting in a size that is approximately one-ninth of the original version of FP16. This model supports 262,000 token contexts, image and text input, and agent tool calls, and it offers its weights under a Apache 2.0 license. Compared to the first generation of this model, which also had 27 billion parameters and was released two months ago, the new version has hardly changed in size, but the underlying model has been replaced from Qwen3.6 to Qwen3.8. Self-tests indicate that the retention rate of comprehensive benchmark scores has increased from about 95% to 98.2%, with math and code scores remaining close to those of the original version. The official team also demonstrated that this model can be directly used for Cline programming and computer tasks on RTX 5090.