PrismML is a Caltech spin-off working on extreme compression of large language models for on-device inference, backed by Khosla Ventures. The company raised a $16.25 million seed round in early 2026 (Source: theinformation.com).
Qwen 3.6 iPhone demonstration
On July 9, 2026, PrismML announced it had compressed Alibaba's 27-billion-parameter Qwen 3.6 model from roughly 54 GB to under 4 GB and ran it with all parameters activated on an iPhone 17 Pro — reported as a record for model size running on a phone (Source: news.futunn.com; theinformation.com). The claim is company-reported and had not been independently verified at the time of coverage.
Apple has held discussions with the company, in the context of Apple's search for on-device AI capability (Source: theinformation.com). By July 18, 2026, Apple was reported to be testing PrismML technology to run large models directly on iPhones (Source: exponentialview.co).
Bonsai 27B
On July 18, 2026, PrismML publicly released Bonsai 27B, a compressed multimodal model based on Qwen3.6 27B, under the Apache 2.0 license. The release ships a 5.9 GB ternary version for laptops and a 3.9 GB 1-bit version for iPhones, which the company reports retain roughly 95% and 90% of full-precision performance respectively (Source: prismml.com; exponentialview.co). The release follows the company's July 9 iPhone demonstration and moves its compression work from demonstration to an openly licensed model.
Relationships
- related: Apple — reported discussions on on-device AI
- related: Alibaba / Qwen Team — compressed the Qwen 3.6 model for the demonstration
- related: Edge AI / Private Physical AI, Distillation — model-compression and edge-inference context