Prism ML shrinks Qwen3.8 into 5.9GB Bonsai 2 27B

Prism ML Inc.'s Bonsai 2 27B arrives as a second-generation attempt to squeeze a capable model onto consumer hardware. The company used ternary compression, simplifying the model's weights from 16 bits down to three bits represented by +1, 0, and -1, which lets it store information in a much smaller memory footprint. The article notes that other compression methods called quantization usually strip away accuracy, knowledge and other capabilities, and that Qwen3.8's minimal memory footprint is 9.4 gigabytes.

On benchmarks, Bonsai 2 stayed close to Qwen3.8 within three points on agentic and tool calling, at 77.6 and 79.8 respectively. It scored 81.6 and 82.2 for coding and 82.7 and 81.3 for knowledge and reasoning. The company said the model consumes extremely low power per token at 0.714 megawatt-hours, making it 40% more energy-efficient than other 8B models running at full precision.

The appeal of ultra-small models is that they let users run AI locally without sending inference to the cloud, which can introduce delays or expose sensitive information to third parties. Whether Bonsai 2 27B delivers on that promise for everyday users and enterprises may depend on how well its benchmark parity translates to real tasks, and on whether the local hardware it targets can handle the workloads users actually care about.