PrismML
5 Intel
PrismML is an AI startup that emerged from Caltech, founded by professor Babak Hassibi, whose research on the mathematical foundations of neural-network compression underpins the company's technology. The company builds 1-bit and ternary-weight compressed models, including its open-source Bonsai family, and drew attention in 2026 for compressing a 27-billion-parameter model to run on an iPhone. It raised a $16.25 million seed round backed by Khosla Ventures, and Apple has reportedly been evaluating its technology for on-device AI.
Is this you? Sign in with X to claim this profile.
Intel
Prism ML released Ternary Bonsai 2 27B, a ternary-quantized model built on Qwen3.8 27B. It fits in 5.9 GB, keeps 98.2% of full-precision aggregate benchmark scores by the vendor's account, and is Apache 2.0 licensed. Demos show Cline coding and local computer use on an RTX 5090.
PrismML tests its ternary Bonsai 2 27B on 2026 IMO problems against full-precision Qwen and a Gemma 12B QAT model, and shows it fixing its own single-file game and desktop builds across feedback rounds. Results are self-reported, with candid notes on multi-turn limits.
PrismML releases Ternary Bonsai 2 at 2B, 4B and 9B under Apache 2.0, built on Qwen3.5. Footprints run 0.43 to 2GB, with vendor-reported retention across 19 benchmarks, tool calling, vision, and up to 34 tok/s for the 9B on an iPhone.
PrismML and Qualcomm optimized a 1-bit Bonsai vision-language model for the Snapdragon Hexagon NPU, claiming 4x lower memory use and 2x faster token generation for equivalent quality, enabling local inference on AR1 Gen 1 smart glasses. Benchmarks were not included in the thread.
PrismML's Bonsai 2 27B uses ternary weights (1.76 effective bits) to shrink a 27B multimodal model to a 5.9GB footprint, retaining 98.2% of the full-precision Qwen3.8 27B's benchmark performance across coding, reasoning, and agentic tool use, with a 262K context window under Apache 2.0.