Apple in Talks with PrismML as China Approves Qwen for Apple Intelligence

Apple is advancing its AI capabilities on two fronts: talks with startup PrismML to run large models on-device and regulatory approval in China to integrate Alibaba's Qwen model into Apple Intelligence.

On-Device AI Compression

PrismML, a Caltech spinout backed by Khosla Ventures, released compressed versions of Alibaba's open-source Qwen model on Tuesday. The company reduced the model from roughly 54GB to under 4GB, allowing its 27 billion parameters to run on an iPhone 15 or newer Source: CNBC. PrismML CEO Babak Hassibi told CNBC that Apple and other companies are evaluating the technology, calling the discussions "very early" but noting "things are progressing nicely" Source: The Next Web. The compression technique reduces each value from 16 bits to one or three possible values, cutting memory use by 10–15 times and speeding up responses by 6–8 times, while sacrificing about 5–10% performance Source: TechCrunch.

China Approval with Alibaba

China's Cyberspace Administration approved Apple Intelligence for launch in the country, integrating Alibaba's Qwen model across iOS, iPadOS, macOS, and visionOS Source: TechCrunch. Alibaba confirmed the partnership, stating Qwen will be "integrated into Apple Intelligence experiences" for text and image understanding and generation Source: The Next Web. The approval ends a lengthy regulatory process and follows Apple's earlier failed attempts with Baidu, DeepSeek, and ByteDance Source: TechCrunch. Alibaba's US-listed shares rose nearly 4% on the news Source: CNBC.

Considerations

• Privacy vs. Performance: On-device AI supports Apple's privacy pitch but may face trade-offs in factual recall and battery life, according to analysts Source: The Next Web.
• Geopolitical Signals: The Qwen deal comes amid US-China tech rivalry; Alibaba recently banned employees from using Anthropic's Claude Source: The Next Web.

Analysts caution that real-world performance across millions of devices remains to be proven, with power consumption as a key question Source: The Next Web.