PrismML
PrismML is a Caltech spinout building ultra-dense 1-bit and ternary large language models for on-device AI inference.
Operating headquarters
Pasadena, CA, USFounded
2025PrismML is a Caltech spinout building ultra-dense 1-bit and ternary large language models for on-device AI inference.
prismml.comCrawled Sep 27, 2026
Explore PrismML research, product updates, and technical deep dives on 1-bit and ternary Bonsai models, on-device AI, and efficient inference.
Today we're announcing Bonsai 27B, our multimodal flagship: ternary at 5.9GB for laptops, 1-bit at 3.9GB for an iPhone 17 Pro, with a 262K-token context.
Read PrismML news, model launches, and media coverage on Bonsai 1-bit and ternary AI models designed for phones, laptops, and efficient computing.
Explore open AI/ML, platform, product, developer relations, and recruiting roles at PrismML in Pasadena or San Francisco.
Meet the PrismML team advancing intelligence density with smaller, high-performance AI models for smartphones, robots, IoT devices, and efficient data centers.
PrismML builds 1-bit and ternary Bonsai models that pack advanced reasoning, tool use, and multimodal AI into phones, laptops, and efficient data centers.
Read PrismML’s Privacy Policy to learn how it collects, uses, discloses, and protects information from website visitors and service users.
Read PrismML’s Terms & Conditions for using its website, content, features, and services, including eligibility, acceptable use, and legal terms.
PrismML has adapted its 1-bit Bonsai language model for smart glasses powered by Qualcomm’s Snapdragon AR1 Gen 1. The 2-billion-parameter model is tuned for vision and language, enabling wearers to ask about what they see in real time, and Qualcomm demonstrated it at Snapdragon Summit. PrismML says its model compression retains nearly all performance on standard benchmarks, but no smart glasses using it have been announced.
Sep 2026 · TechCrunchNews article
Prism ML Inc. launches Bonsai 2 27B, a multimodal model compressed from Qwen3.8 27B to about 5.9 GB using ternary weights while retaining about 98.2% of its capabilities. It runs locally on Nvidia GPUs and Apple devices through CUDA and MLX, and the company reports benchmark scores close to the base model across agentic tasks, coding, knowledge, and reasoning. Model weights are available under Apache 2.0 licenses.
Sep 2026 · SiliconAngleNews article
PrismML releases Bonsai 2 27B, a compressed version of Alibaba’s Qwen3.8 27B that fits in 5.9 GB while retaining 98% of its aggregate benchmark scores. Founded by Caltech researchers and led by professor Babak Hassibi, the startup uses ternary weights and has raised a $22.25 million seed round, with backing from Khosla Ventures, Cerberus Capital, and Caltech. It plans to apply the technique to models with several hundred billion parameters.
Sep 2026 · TechCrunchNews article
GitHubRepository owner