Apple Intelligence reaches China on Qwen — dev testing opens Nov 4
Beijing's sign-off puts Alibaba's model behind the iPhone's built-in assistant, with Baidu filtering content; plus LeCun's 42x-fewer-samples JEPA proof.
Copy markdown
China clears Apple Intelligence — on Alibaba's Qwen
Chinese regulators approved Apple Intelligence for the mainland with Alibaba's Qwen as the underlying model and Baidu handling content compliance. Developer testing opens Nov 4, with the public rollout arriving via iOS 27.3 in December across iOS, iPadOS, macOS and visionOS. If you ship apps into China, the OS-level assistant your users touch now runs on Qwen — not Apple's own models — so plan prompts and behavior around Qwen's quirks.
JEPAs reportedly match quality on 42x fewer samples
Balestriero and LeCun posted a proof-plus-experiments result arguing joint-embedding predictive architectures (JEPAs) learn good representations with reportedly ~42x fewer training samples than next-token prediction. It's a paper, not a shippable model — but it's LeCun's clearest swing yet at next-token orthodoxy, and one to watch if you pre-train or distill your own small models on tight data or compute budgets.