AI startup PrismML has released Bonsai 2 27B, a compact version of Alibaba’s Qwen3.8 27B model. The model, with 27 billion parameters, was packed into 5.9 GB. According to PrismML tests, it retains 98.2% of the results of the original.
AI startup PrismML has released Bonsai 2 27B, a compact version of Alibaba’s Qwen3.8 27B model. The model, with 27 billion parameters, was packed into 5.9 GB. According to PrismML tests, it retains 98.2% of the results of the original.
This is reported by TechCrunch.
A full-size Qwen3.8 27B in 16-bit format requires about 54 GB of memory. PrismML has reduced this amount by more than nine times. To do this, the company used ternary weights. Weights are numerical values in which the model stores what it has “learned” during training. Instead of large numbers, PrismML represents them with only three possible values: −1, 0, or +1. Thanks to this, the model takes up much less space.
PrismML also claims that Bonsai 2 27B performs 98.2% of the original Qwen3.8 27B in a combination of tests for logical reasoning, mathematics, programming, following instructions, working with images, and using tools by AI agents. However, these are PrismML's own benchmarks, not independent testing.
The model supports contexts of up to 262,000 tokens, can work with text and images, and is available under the Apache 2.0 open license. It can be run via CUDA, as well as on devices with Apple Silicon. Now the startup wants to take on even larger models — with several hundred billion parameters. PrismML CEO Babak Hassibi told TechCrunch that the company plans to show the first such models in the coming months.
Previously, dev.ua wrote about Google's TurboQuant, a technology that should reduce the amount of memory required to run large language models without a noticeable loss of accuracy.
