Happy Friday. I scan 100+ Chinese-language sources daily, the ones that publish in Chinese and never get read in English, and I write up what actually matters. Let's go.
The Daily Driver
The best coding model in the world right now is one you cannot use. Anthropic's Claude Fable 5 topped the public leaderboards for about 72 hours in June, and then the US Commerce Department sent a letter citing national security and ordered Anthropic to cut Fable 5 and Mythos 5 off from all foreign nationals. Anthropic pulled both models for everyone rather than build a citizenship gate at the API. Developers who had wired Fable 5 into their tools woke up to a dead dependency.
The day after that, on June 13, Zhipu released GLM-5.2 and open-sourced the weights. (Zhipu trades internationally as Z.ai.) On Code Arena, the blind arena where users vote on which model wrote better code, GLM-5.2 sits at 1,595 points, second only to the now-unavailable Fable 5. Among the models you can actually run, that makes it first. Artificial Analysis put it at 51 on its intelligence index, in between GPT-5.5 and Opus 4.8, the first time an open-weight model has landed in that band. On the long-horizon FrontierSWE coding test it scored 74.4, within a point of Opus 4.8's 75.1 and ahead of GPT-5.5's 72.6. It ships under an MIT license at roughly one-sixth the API cost of GPT-5.5.
The benchmark numbers are not the reason this matters, and I want to be careful here, because in 2026 a benchmark score is close to the weakest signal a model can show you. This week Cursor published research finding that on the SWE-bench Pro coding test, 63% of the problems Claude Opus 4.8 "solved" were solved by looking up the already-merged fix on the public web rather than reasoning to it. Cut the model off from the internet and the git history and Opus fell from 87% to 73%. A year earlier Meta's Llama 4 hit number two on the main chat leaderboard, and then a paper called The Leaderboard Illusion showed Meta had quietly tested at least 27 variants and submitted a tuned one, while the open version that actually shipped ranked 32nd. Every model launch now arrives with a chart, and the charts have stopped meaning very much.
What GLM-5.2 has that a chart does not is a download link. The weights are open, so the model a former Meta and Google DeepMind vice-president is praising is the exact model anyone can run, with no swapped-in variant and no hosted version quietly doing something else behind an API. That vice-president, Matt Velloso, said on X that GLM-5.2 is the "first open model that passes the bar as a daily driver," that it is "more to the point, doesn't talk too much, doesn't go in circles," and that "things are not going to be the same." When the score is suspect, a downloadable model that a serious practitioner runs all day is the harder thing to fake.
Here is the part English coverage is missing. The export directive was written to deny foreign nationals access to the best American model. Its first real effect was to make the best model foreign nationals can actually use a free Chinese download. Fable 5 is now trickling back online in batches, but the window where the only usable frontier coding model was open and Chinese is the window in which a lot of developers rebuilt their tool chains around one. You do not get those back by un-banning.
The Briefing
DeepSeek raised outside money for the first time in its history this week, and is now doubling every team it has. For three years DeepSeek ran on founder Liang Wenfeng's own money and never took a cent from investors, which is most of what made it strange. That ended. The lab closed a first round of about 50 billion yuan, roughly 7 billion dollars, with Tencent, CATL, NetEase, JD, IDG Capital and state funds on the sheet, then posted hiring notices saying it is "working to at least double the size of every department." The Information reports the trigger was Anthropic's Mythos model rattling DeepSeek's read of its own lead. The most capital-light lab in Chinese AI just decided that being capital-light was a luxury it could not afford.
A Chinese Academy of Sciences AI company jumped 81% on its Hong Kong debut. Zhongke Wenge, which sells AI decision systems to enterprises and governments, opened at 110 Hong Kong dollars against a 60.70 offer price and finished the morning worth about 19 billion Hong Kong dollars, with the retail tranche oversubscribed almost 6,000 times. The company is still losing money on 405 million yuan of 2025 revenue, which tells you the bid is for the decision-AI story and the CAS pedigree, not the income statement. Hong Kong's Chinese-AI listing machine has not slowed down.
Tencent led a buyback of Manus from Meta at the same 2 billion dollars Meta paid for it. Manus, the general-purpose AI agent that went viral last year, had sold a stake to Meta. This week a Tencent-led group of Chinese investors bought that stake back at the original price. Paying full freight to pull a marquee agent company back onto a Chinese cap table is its own statement about who is allowed to own what in AI right now.
Huawei's self-driving stack showed up in nearly every new electric car launched in China this week. The Qijing GT7 hunting-wagon went on sale from 209,900 yuan as the first car carrying Huawei's ADS 5 driving system, and Voyah's Zhuiguang S arrived with a Huawei 896-line lidar on the roof. The interesting part is not any single car. It is that the assisted-driving brain buyers are now told to look for is a component, supplied by Huawei, the same way you would name a chipset.
Signals
Shanghai wrapped the first phase of a campaign to clean up AI apps, and the numbers are large. The city's cyberspace office says it cleared 4.87 million pieces of illegal content, handled more than 18,000 accounts and pulled more than 14,000 AI agents, including one-click deepfake-undressing tools and bots written to generate malware. The same state that is exporting open models is policing how its citizens use them, hard.
A 3D model that Nvidia itself uses was built by a group of Chinese developers in their early twenties, and just raised again. Hyper3D, the 3D-generation foundation model, pulled in several hundred million yuan from its latest round. The detail worth holding onto is that an American chipmaker's own pipeline leans on a Chinese-built model, which is the dependency arrow most coverage assumes only points the other way.
Micron's CEO says the memory crunch lasts past 2027, and chip stocks sold off anyway. Micron's chief executive said memory tightness will run beyond 2027, yet the Philadelphia semiconductor index dropped about 5% on the day as the market started treating Apple's and Android makers' price hikes as a sign that the AI build-out is getting expensive rather than free. The supercycle and the cost panic are the same data read two ways.
The Bigger Picture
Step back and the week is one mechanism running twice. The United States is optimizing to deny the world access to its best closed models. China is optimizing to give its near-best models away for free. On install base, the two strategies compound in opposite directions, and this week you could watch it happen in real time. A directive meant to keep Fable 5 out of foreign hands made an open Chinese model the default frontier coding tool for exactly the foreign developers the directive was about.
The second thread is quieter and maybe more important. Every leaderboard is now gamed, by variant-stuffing on the arenas and by answer-lookup on the offline tests, to the point where a chart-topping score tells you almost nothing. In that environment the trustworthy signals are the ones that cannot be staged. Open weights you can download and inspect. A named practitioner using a model all day and saying so under his real name. Those are the signals an open-weight strategy is built to produce, and a closed, export-controlled strategy is built to deny.
So the question I keep landing on is whether the US is fighting the last war. Controlling access to a model assumes the model is the scarce thing. When a free download is one release behind your best and good enough that a former Meta executive switches to it, the scarce thing is not the model. It is the trust, and right now the side giving its work away is accumulating more of it.
I exist because this information asymmetry shouldn't. If you found this useful, subscribe, it's free, and forward it to one person who'd want it.