The Day in AI

• Xiaomi released open weights for Xiaomi-OCR-0 on 30 October, a 0.8-billion-parameter document model scoring 96.83 on OmniDocBench v1.6 by its own measure.
• Xiaomi says the same small model reads documents, pulls out key fields and answers questions about them, and already runs in its Super XiaoAi phone assistant.
• Apple said on 15 October that Apple Intelligence, running on Alibaba's Qwen models, has Chinese regulatory approval, with a beta due in late October.
• Anthropic says an attacker used access to Cursor's marketplace to publish a tampered copy of its Claude Code extension, reportedly live for nearly 12 hours.
• Google joined OpenAI and Anthropic in backing a new standard that lets websites opt out of AI scraping, The Verge reports.

Xiaomi's model does three jobs with one set of weights. It parses documents, extracts key information and answers visual questions about text in images. It is built on Qwen3.5-0.8B-Base and trained on about 170 million samples in three stages.

The last stage uses reinforcement learning (RL) with verifiable rewards. The model's output is scored automatically against a known answer, for example by the structural edit distance between its reconstructed layout and the reference, or by how well its tables match. The model then learns from those scores directly and needs no human ratings.

All figures are company-reported. The 96.83 on OmniDocBench v1.6 puts it second among all models on that benchmark, 0.04 points behind the leader, and the reports don't name that leader. Xiaomi gives 95.24 on Real5-OmniDocBench, which uses simulated scans, curled paper and photos of screens, and 87.94 on Wild-OmniDocBench. On formulas it reports 97.84 on UniMER-Test, and on table structure 92.27 on PubTabNet.

On question answering, Xiaomi averages five benchmarks including DocVQA, ChartQA and OCRBench and gets 83.2. That beats Tencent's HunyuanOCR-1.5 at 76.5 and the 8-billion-parameter MiniCPM-V-4.5 at 82.6, and trails Qwen3.5-4B at 84.9. A model this small fits on a phone, and Xiaomi says it already runs in Super XiaoAi.

Apple said on 15 October that Apple Intelligence has regulatory approval in mainland China, where it has been missing since launching elsewhere. The China version uses Alibaba's Qwen and Tongyi Wanxiang models, and Baidu reportedly handles search and some Siri functions. Tim Cook, Apple's chief executive, confirmed a beta for late October.

The first batch has around 18 features, according to an Apple statement reported by the South China Morning Post, including live translation, writing tools and photo cleanup. The redesigned Siri is not part of it. A wider launch is reported for December in iOS 27.3, and Apple hasn't confirmed that date.

• Anthropic says a malicious actor used access to Cursor's marketplace to publish a tampered Claude Code extension that Anthropic didn't build, reportedly live for nearly 12 hours.
• Google joined OpenAI, Anthropic and Microsoft in backing a new opt-out standard for AI scraping, according to The Verge.
• DeepSeek is reportedly testing a dedicated cybersecurity model called Cyber, according to Pandaily, which found references in the company's website code.
• Tsinghua University and Zhipu describe HarnessPro, a method that lets agents rewrite their own harness, reportedly cutting task costs by 26%.
• Sebastian Thrun, co-founder of Google X, raised $130 million for Dulo, an AI engineering startup, SiliconANGLE reports.
• Salesforce is reportedly buying the defense AI startup Rune for $375 million, according to TechCrunch.
• Anthropic is expected to get a high "junk" credit rating after its initial public offering, the Financial Times reports.