Apple in talks with PrismML on iPhone AI model compression

PrismML, a Caltech spinout, has compressed Alibaba's 27-billion-parameter Qwen model from 54 GB to under 4 GB, allowing all parameters to run fully active on an iPhone 15 or newer. The compression technique uses 1-bit weights (on/off states) and ternary variants instead of standard 16-bit floating-point format, achieving a 90%-plus compression ratio. CEO Babak Hassibi told CNBC that Apple and other firms have been evaluating the technology and the discussions are "progressing nicely"; Hassibi said PrismML's compressed models use 10–15x less memory, run 6–8x faster, and consume 3–6x less energy than conventional versions, though with small accuracy trade-offs.

This addresses a key constraint in Apple's on-device AI strategy. Apple's current largest on-device model is AFM 3 Core Advanced with 20 billion parameters using sparse architecture (only 1–4 billion active at once). PrismML's approach keeps all 27 billion parameters active simultaneously, enabling more complex reasoning, code generation, and agentic tasks. The timing aligns with Apple's June WWDC overhaul of Siri, which currently relies on Google's Gemini for advanced features.

Architects watching on-device inference should note the implications for cloud vs. edge economics. If PrismML's claims hold in production, shrinking AI models could shift compute from datacenters to phones, reducing reliance on cloud licensing and lowering latency for private features. PrismML raised $16.25 million seed funding from Khosla Ventures and plans to compress larger frontier models. However, independent verification of performance claims remains critical—analysts flagged battery consumption during continuous tasks and real-world reliability at scale as open questions.