While US desks went dark for the Fourth, China's AI labs shipped. Meituan — the Beijing delivery giant better known for bringing you dinner — open-sourced LongCat-2.0, a 1.6-trillion-parameter model with a one-million-token context window, weights posted to GitHub and Hugging Face under a permissive MIT-style license. It landed days after startup Z.ai released GLM-5.2, an open-weights model that beats GPT-5.5 on multiple agentic coding benchmarks at roughly a sixth of the price.
The LongCat claim that matters isn't the parameter count. Meituan says the model was trained end-to-end on a cluster of 50,000 domestic chips — Huawei's Ascend line — with no US hardware in the loop. If that holds, it's the first model of this scale trained entirely outside Nvidia's ecosystem, and a direct answer to Washington's export controls. The company pegs performance as comparable to Google's Gemini 3.1 Pro, released in February. Caveat where it's due: those are Meituan's own numbers, and some observers have flagged that the technical artifacts behind the domestic-chip claim are still thin.
GLM-5.2 needs fewer caveats, because the weights are public and the results are being kicked hard. The 744-billion-parameter model scores 62.1 on SWE-bench Pro against GPT-5.5's 58.6, and 74.4 on FrontierSWE — ahead of GPT-5.5's 72.6 and within a point of Claude Opus 4.8's 75.1. Z.ai charges $1.40 per million input tokens and $4.40 per million output; third-party hosts go lower. That is frontier-adjacent coding intelligence at commodity prices, license attached, no permission required.
The price war just went international
Two weeks ago, Sonnet 5 repriced autonomous work from above. This weekend the pressure arrived from below and abroad at once: models you can download free, run on your own hardware, and fine-tune without asking anyone. And it lands while the US and its allies pour money into the other side of the equation — South Korea's $880 billion chip mobilization and Anthropic's hunt for its own silicon are bets that compute advantage stays decisive.
Our take: Export controls were supposed to be a wall. LongCat-2.0 — if the training claim survives scrutiny — says the wall leaks; GLM-5.2 says it may not matter. The real casualty is the mid-market: when good-enough coding intelligence is free and the top tier is a benchmark point away, closed labs are pushed to defend a narrowing summit while volume drains to open weights. Watch enterprise compliance teams, not benchmark charts — they, not developers, will decide whether Chinese open weights are deployable in the West.
What to watch
- Independent replication of LongCat's domestic-chip training claim. It's the boldest assertion of the week, and it's still self-reported.
- Western hosting. MIT licensing means US clouds could serve both models tomorrow. Whether they do is a policy question, not a technical one.
- Pricing responses from OpenAI and Anthropic — with White House release standards possibly days away, closed labs face new compliance costs just as free alternatives multiply.