PrismML releases Bonsai 2 27B with significantly reduced memory footprint
AI lab PrismML has launched Bonsai 2 27B, a version of its model that achieves 98.2% of the performance of the Qwen3.8 27B base model while using less than one-ninth of the memory.
AI laboratory PrismML has introduced a new version of its AI model, Bonsai 2 27B, featuring a three-value quantization that significantly reduces memory requirements. The model upgrades the base architecture to Qwen3.8 27B, enhancing performance in inference, programming, vision, and long-context reasoning.
Bonsai 2 27B achieves 98.2% of the performance of its Qwen3.8 27B counterpart while consuming less than one-ninth of the memory. This efficiency allows the model to handle demanding tasks locally, such as running programming agents, managing computer workflows, parsing private documents, and processing multimodal data, reducing the need for cloud API calls.
The model utilizes FP16 grouped quantization, resulting in an effective bit precision of 1.76 per weight. It has a total size of 5.9GB and supports a 262K context window. Its energy efficiency on a GeForce RTX 4090 is reported to be 0.714 mWh per token, a 40% improvement over a full-precision 8B model.
Bonsai 2 27B can be run on NVIDIA GPUs via CUDA and on Apple devices using MLX. The model's weights are released under the Apache 2.0 license, enabling wider adoption and further development.