PrismML Releases Compact Bonsai 2 27B AI Model

PrismML has released Bonsai 2 27B, a new AI model based on Qwen3.8 27B. According to the company, it can handle tasks involving analysis and reasoning, coding, image processing, and multi-step actions while using significantly less memory than the full-size version.

The compressed version of Bonsai 2 27B requires just 5.9 GB of memory — more than 9x less than the full-precision model. At the same time, it retains 98.2% of its overall benchmark performance. This is an improvement over the first Bonsai 27B, which retained 95% of the full-size model’s performance.

PrismML tested the new model across 20 benchmarks covering reasoning, mathematics, coding, instruction following, image processing, and tool use. It achieved an aggregate score of 83.9 points, compared with 85.4 for Qwen3.8 27B. On an NVIDIA GeForce RTX 5090, Bonsai 2 27B can run at speeds of up to 143 tokens per second.

The model has 27.8 billion parameters and was trained on Google v5 TPUs. It is designed to run efficiently on consumer CPUs and GPUs and can be used for local AI assistants, multimedia applications, private data workflows, and tasks that require AI to carry out multiple actions in sequence. The compact version is already available for free under the Apache 2.0 license.