Tip Of AI
Discover what you can do with AI
PrismML Bonsai is a suite of ultra-dense AI models designed to concentrate advanced reasoning, multimodal understanding, and efficient tool use into devices with limited resources. The models employ both 1-bit and ternary quantization methods that shrink memory footprints by up to nine times while reducing energy consumption by five times and boosting inference speed by eight times. These optimizations maintain over 98 percent of full-precision performance on flagship models like Bonsai 2 27B, which runs in just 5.9 GB of RAM and supports long-horizon tool use and agentic coding. PrismML prioritizes intelligence density, defined as the negative logarithm of error rate divided by model size, rather than sheer parameter count. Supported by Caltech research, the Bonsai series spans 1.7B, 4B, 8B, and 27B parameter sizes in both 1-bit and ternary formats. These models are built on the Qwen3.8 architecture and represent a major step toward closing the quality gap with full-precision networks. Users can download prequantized models or try a live demo, and detailed benchmarks on throughput, energy consumption, and error rates are provided in the company whitepaper.