정규

Qwen3.8-27B reportedly went from 55.6GB to just 5.9GB, while retaining 98.2% of its aggregate benchmark performance. That’s an approximately 89.4% reduction in model size. The result comes from PrismML, the team behind Bonsai, which has introduced its second-generation compression system. Here are the reported numbers: Original Qwen3.8-27B: 55.6GB Bonsai 2 compressed version: 5.9GB Model size reduction: 89.4% Aggregate benchmark performance retained: 98.2% Previous Bonsai 27B version: approximately 95% retention Evaluation coverage: 20 benchmarks The interesting part isn’t just how much smaller the model becomes. It’s how much capability reportedly survives such aggressive compression. How does Bonsai 2 achieve this? The underlying technique is ternary quantization. Most conventional model weights use numerical representations with multiple possible values. Depending on the format, each weight may use 16, 8 or 4 bits. Ternary quantization takes a more aggressive approach. Instead of preserving a large range of numerical values, weights are represented using three possible states, typically: -1 0 +1 Scaling factors help recover useful numerical ranges from these compact representations. Because each weight has only three possible states, the storage requirement can be dramatically reduced compared with conventional higher-precision formats. But there’s a difficult trade-off. A model’s weights encode the information that allows it to reason, follow instructions, write code and solve problems. Compressing them too aggressively can destroy distinctions that matter to those capabilities. The challenge is therefore not simply reducing the number of bits. It’s finding a way to remove numerical precision without removing too much useful information. What improved in Bonsai 2? PrismML reports better capability retention than its previous 27B compression, moving from approximately 95% to 98.2% aggregate benchmark performance. The team also highlights improvements in: Coding agents Multimodal reasoning Long-horizon tool use The agentic side is particularly interesting. Consider a coding agent working on a real project. It may need to inspect a codebase, understand dependencies, modify multiple files, run tests, interpret errors and repeat the process until the requested feature works. Preserving that kind of behaviour after reducing the model to 5.9GB is more interesting than simply maintaining accuracy on short question-and-answer tests. An agent needs to maintain context, make decisions across multiple steps and recover when something goes wrong. If those abilities survive compression, the practical value goes well beyond having a smaller model file. Why this matters for local AI A 5.9GB model is considerably easier to distribute, store and experiment with than a 55.6GB version. It could make a 27B model more accessible to people using consumer GPUs, laptops and systems with limited memory. There is one important distinction, though: 5.9GB is the reported compressed model size, not necessarily the total VRAM required to run it. Inference still needs memory for the KV cache, activations, temporary buffers and runtime overhead. Context length and concurrency can increase those requirements. The 98.2% figure also represents aggregate performance across 20 benchmarks. It does not mean the model performs identically to the original on every task. Some capabilities may retain more performance than others. Independent testing will be important for understanding where the remaining performance loss occurs, particularly on coding agents, complex reasoning and long-running tool workflows. Still, moving from approximately 95% retention in the previous Bonsai version to 98.2% with Bonsai 2 is a notable reported improvement. The real test of compression isn’t how many gigabytes you remove. It’s how much useful capability remains after you remove them.