Raphael🔶

Qwen3.8-27B has reportedly been compressed from 55.6GB to just 5.9GB, while retaining 98.2% of its aggregate benchmark performance. That’s an 89.4% reduction in model size. The work comes from PrismML, which has released Bonsai 2, the second generation of its model compression system. What makes this result interesting is that the reduction in storage hasn’t been accompanied by a comparable collapse in the reported benchmark results. The numbers behind Bonsai 2 Original Qwen3.8-27B checkpoint: 55.6GB Bonsai 2 checkpoint: 5.9GB Storage reduction: 89.4% Aggregate benchmark performance retained: 98.2% Evaluation coverage: 20 benchmarks Previous Bonsai 27B version: approximately 95% retention The improvement over the first generation is worth noting. PrismML reports that performance retention increased from roughly 95% to 98.2%, despite maintaining an aggressive compression target. That suggests the team has improved how it removes numerical information from the model without disrupting as much of its useful behaviour. How does Bonsai 2 achieve this? The central technique is ternary quantisation. Conventional neural networks store weights using numerical formats such as FP16, INT8 or 4-bit representations. These formats support different levels of numerical precision, with corresponding storage requirements. Ternary quantisation takes a more aggressive approach. It represents model weights using just three possible values: -1 0 +1 Scaling factors help preserve information about the magnitude of the original weights. Instead of storing every weight as a conventional high-precision number, the system uses a much smaller representation and scaling information to recover useful numerical ranges. This dramatically reduces the storage required for the model. However, there is a difficult problem with aggressive quantisation. Model weights encode relationships that influence how a model reasons, writes code, follows instructions and processes information. Removing too much numerical detail can distort those relationships and damage the capabilities developers depend on. The objective isn’t simply to reduce the number of bits. It’s to find a representation that removes unnecessary numerical precision while preserving as much useful model behaviour as possible. What improved over Bonsai 1? The reported increase from approximately 95% to 98.2% aggregate performance retention is one part of the Bonsai 2 update. PrismML also highlights improvements in three areas that are particularly relevant to modern AI applications. 1. Coding agents Coding agents need more than the ability to generate a function from a prompt. They must understand existing repositories, navigate dependencies, modify multiple files, execute tests, interpret errors and revise their implementations. A model can perform well on isolated coding questions but still struggle when asked to complete an extended development task. Preserving performance across these workflows is therefore an important test of compression quality. 2. Multimodal reasoning Multimodal models work with different forms of information rather than relying exclusively on text. Depending on the task, this can involve interpreting images alongside written instructions, extracting information from visual inputs and reasoning across different types of content. Compression needs to preserve the relationships required to process these inputs, not just the model’s ability to answer ordinary text questions. 3. Long-horizon tool use This is particularly relevant to autonomous agents. A long-horizon workflow may involve several connected decisions: Inspecting information. Choosing a tool. Executing an action. Reviewing the result. Identifying mistakes. Adjusting the next action. Errors can accumulate across these steps. A model that loses consistency during a long workflow may produce a poor final result even if it performs well on individual questions.