Bonsai Demo

Bonsai is the family of compressed language models from PrismML, and the demo is the hosted way to try them before downloading. PrismML builds what it calls ultra dense intelligence: models quantized to ternary or one bit weights so that capable reasoning fits on phones, laptops and edge hardware instead of a datacenter. The current flagship, Ternary Bonsai 2 27B, is built on a Qwen 27B base and weighs 5.9 GB, which PrismML says is a ninefold smaller footprint than its full precision counterpart while retaining 98.2 percent of its benchmark performance. The models suit developers and device makers who want local inference for agentic coding, multimodal reasoning and long horizon tool use without sending data to a cloud service. Alongside the 27B model the lineup includes Bonsai 8B, 4B and 1.7B in both ternary and one bit variants, plus Bonsai Image. PrismML publishes benchmark comparisons, throughput in tokens per second across hardware platforms and energy use in milliwatt hours per token, with the method set out in a whitepaper. Models can be downloaded directly, tried in the web demo, or used through Bonsai Studio. PrismML does not publish a pricing page. The site offers model downloads and the demo without listed fees, and partnership or commercial enquiries go through a contact form. PrismML grew out of research at Caltech, and its headline claims for the flagship model are nine times less memory, eight times faster inference and five times less energy than a standard 16 bit model, figures the company attributes to its own evaluation.