PrismML reportedly compressed Alibaba's Qwen 3.6 model to run locally on an iPhone 17 Pro. Apple's interest signals a shift toward on-device large models, bypassing cloud dependencies for privacy and speed.
Key facts
- Qwen 3.6 has 36 billion parameters.
- iPhone 17 Pro uses a 3nm A19 Pro chip.
- PrismML's compression preserves most model accuracy.
- Apple has not confirmed a partnership or timeline.
- On-device inference eliminates cloud latency and privacy risks.
PrismML, a startup specializing in model compression, has reportedly achieved a significant milestone: shrinking Alibaba's Qwen 3.6 — a 36-billion-parameter large language model — to run entirely on an iPhone 17 Pro. According to TrendForce The compression method reportedly preserves most of the model's accuracy while reducing memory footprint and inference latency to levels suitable for a mobile device. The iPhone 17 Pro's A19 Pro chip, built on a 3nm process, provides the neural engine horsepower needed for real-time inference, though exact token-per-second rates were not disclosed.
Why Apple Cares
Apple has long pursued on-device AI to differentiate its hardware and address privacy concerns. Running a model of Qwen 3.6's scale locally — without cloud round-trips — would allow features like real-time document summarization, code generation, and contextual assistance without sending user data to servers. The move echoes Apple's earlier work with MLX, its own machine learning framework, but PrismML's approach appears to achieve higher compression ratios. Apple has not confirmed a partnership, but the report claims discussions are underway. No financial terms or timeline have been disclosed.
The Compression Arms Race
PrismML is not alone. Google, Meta, and Mistral have all invested in quantization and distillation techniques to shrink models. Google's Gemma 4 2B, for instance, runs on-device but at a fraction of Qwen 3.6's parameter count. The race is now about preserving capability at smaller sizes: a 36B model compressed to run on a phone could outperform a natively small model in reasoning and coding tasks. If PrismML's technique proves production-ready, it could give Apple a lead in on-device intelligence without relying on Google's or OpenAI's cloud APIs.
What to watch
Watch for Apple's WWDC 2027 keynote: if PrismML's technology is integrated into iOS 21, it will likely debut in developer betas. Also track any patent filings by Apple or PrismML covering the compression method.
Source: news.google.com
