PrismML Packs 27B AI Model Into 5.9 GB for Edge Devices

Founded by California Institute of Technology researchers and led by Caltech professor Babak Hassibi, PrismML is pursuing low-bit compression as an alternative to ever-larger cloud models. The approach matters because shrinking capable reasoning systems enough for PCs and premium smartphones could cut memory use, latency and power consumption while keeping private data on-device. The startup has raised a $22.25 million seed round from backers including Khosla Ventures, Cerberus Capital and Caltech.

On Sept. 17, 2026, PrismML released Ternary Bonsai 2 27B, compressing Alibaba’s open-source Qwen3.8 27B from about 54 GB at FP16 to 5.9 GB, a more than ninefold reduction. PrismML said the model scored 83.9 across 20 benchmarks, retaining 98.2% of the full-precision model’s 85.4 aggregate score. The free Apache 2.0 release supports multimodal text-and-image input and a 262,000-token context window, with reported speeds of up to 143 tokens per second on Nvidia’s GeForce RTX 5090.