Overview
- Multiple outlets reported on Wednesday that Apple has met with Caltech spinout PrismML to discuss using the startup’s on-device model compression technology.
- PrismML says it reduced Alibaba’s open-source Qwen 3.6 model from about 54 GB to under 4 GB so the full 27 billion-parameter model can run on an iPhone 17 Pro without apparent performance loss.
- The startup’s approach reportedly keeps all 27 billion parameters active at once, which differs from Apple’s current on-device model that uses a sparse design and activates only 1–4 billion parameters at a time.
- PrismML plans to publish the compressed model as open source on July 14, 2026, a release that independent researchers can use to test the company’s performance and memory claims.
- If the claims hold up, Apple could reduce reliance on cloud models like Google’s Gemini for advanced Siri features, lower server costs, and keep more user data on device, but no partnership or acquisition has been confirmed.