Analysis
Meituan, the Chinese super-app giant best known for food delivery, open-sourced LongCat-2.0 on June 30, 2026 -- a 1.6-trillion-parameter mixture-of-experts (MoE) coding model with a native one-million-token context window, released under the permissive MIT license across GitHub, Hugging Face and Meituan's own platform. The release simultaneously unmasked 'Owl Alpha,' an anonymous stealth model that had quietly commanded global developer charts on OpenRouter for roughly two months. The model that thousands of developers had been routing to without knowing its origin turned out to be Chinese.
The technical headline is the hardware. Meituan says LongCat-2.0 was trained and is served on a 50,000-card domestic compute cluster containing no Nvidia A100s, H100s or AMD MI300X accelerators -- the entire pipeline ran on Chinese-manufactured silicon. As an MoE system the model does not fire all 1.6 trillion parameters at once; it dynamically activates roughly 33 billion to 56 billion per token, keeping inference costs far below what a dense model of that size would demand. That efficiency is what let an under-the-radar model hold the top of OpenRouter's coding rankings on cost-performance.
How we got here matters. Washington has spent three years tightening export controls to deny China the most advanced GPUs, on the theory that compute is the chokepoint that keeps Chinese labs a generation behind. China's answer has been a parallel buildout -- DeepSeek's low-cost training runs, Alibaba's Qwen series, Moonshot's Kimi, and now Meituan -- each chipping away at the premise that frontier capability requires American hardware. LongCat-2.0 is the most pointed rebuttal yet because it claims an end-to-end domestic stack, not just clever optimization on smuggled chips.
“LongCat-2.0 is the most pointed rebuttal yet because it claims an end-to-end domestic stack, not just clever optimization on smuggled chips.”
The competitive landscape is now openly split between two philosophies. On one side, OpenAI and Anthropic increasingly gate their strongest models -- GPT and Claude's most capable tiers -- behind vetting, enterprise contracts and, in some cases, government disclosure. On the other, Chinese labs are flooding the zone with permissively licensed weights anyone can download, fine-tune and embed. An MIT license is the most aggressive posture possible: it lets developers fold LongCat-2.0 directly into closed-source, proprietary commercial tools with no copyleft strings attached, turning a state-adjacent model into free infrastructure for the global developer base.
On the numbers, a 1.6T MoE that leads OpenRouter on real developer usage is not a toy. For comparison, the agentic-coding category Western buyers pay for -- Cursor, GitHub Copilot, Claude Code, Chamath's newly funded 8090 Labs -- monetizes the exact capability LongCat now gives away. If a free, self-hostable model is competitive on coding tasks, the price umbrella over paid frontier coders compresses, and the question shifts from raw capability to distribution, tooling and trust.
For founders, GPs and LPs, the read is twofold. First, open weights from China are becoming a genuine substitute for paid API calls in cost-sensitive, code-heavy workloads -- a real margin threat to application-layer startups built on metered frontier tokens. Second, defensibility is migrating away from the model itself toward proprietary data, distribution and workflow lock-in, exactly the thesis Base44 invoked this week when it launched its own model. The model is increasingly the commodity; the moat is everything around it.
The bear case is real. Claims of a fully domestic 50,000-card training run are hard to independently verify, and benchmark leadership on OpenRouter is not the same as topping rigorous evaluations or winning enterprise trust -- many Western buyers will not run a Chinese state-adjacent model on sensitive code regardless of price or license. Data-governance, security and geopolitical concerns are a structural ceiling on adoption. What to watch: independent benchmarks and red-teaming of LongCat-2.0, whether US policymakers respond to the 'trained without our chips' framing with tighter controls or a strategic rethink, and how Western labs answer an open-source wave that keeps narrowing the gap.
