Meituan — yes, the food delivery app — just trained a 1.6 trillion parameter AI model entirely on Chinese hardware. No Nvidia. No H100s. No smuggled GPUs through Malaysia. Just 50,000 domestically made AI ASICs running in a Chinese data center, crunching through 35 trillion tokens.
The model is called LongCat-2.0, and it's not some academic toy. On SWE-bench Pro — the benchmark that measures whether an AI can actually write production code — it scored 59.5. That's better than Gemini 3.1 Pro and GPT-5.5. Only Claude Opus 4.7 and 4.8 beat it. On SWE-bench Multilingual, it hit 77.3, topping both Gemini and GPT.
The LongCat team didn't exist three years ago. Meituan spun it up in 2023, shipped its first model in late 2025, and now has a trillion-parameter beast that competes with the best Western models on coding tasks. The team claims this is the first trillion-parameter model trained entirely on domestic Chinese compute.
The Chip That Dare Not Speak Its Name
Here's where it gets interesting: Meituan won't say which Chinese chips they used. The announcement mentions "domestically made AI ASICs" but refuses to name the vendor. The obvious candidates are Huawei's Ascend chips or chips from Biren Technology or MetaX. But Meituan's silence suggests either commercial sensitivity or — more likely — a desire not to paint a target on their supplier's back.
Washington has been trying to kneecap China's AI ambitions since 2022 with export controls on advanced semiconductors. The logic was simple: no Nvidia GPUs, no competitive Chinese AI. The US even closed loopholes that let Chinese companies buy downclocked "China-special" versions of Nvidia chips.
LongCat-2.0 is Beijing's answer: We don't need your chips anymore.
What the Benchmarks Actually Show
Before anyone declares the death of Nvidia, let's look at the full picture. LongCat-2.0 wins on coding benchmarks but trails on others:
Where it wins:
- SWE-bench Pro: 59.5 (beats Gemini 3.1 Pro, GPT-5.5)
- SWE-bench Multilingual: 77.3 (beats Gemini, GPT)
Where it loses:
- IFEval: 90.0 (trails Gemini and GPT-5.5)
- IMO-AnswerBench: 81.8 (behind Western models)
- GPQA-diamond: 88.9 (solid but not leading)
The pattern is clear: LongCat-2.0 is a coding specialist. It was trained by Meituan, a company that runs one of the world's largest delivery and logistics platforms. Code generation matters for them — automating backend systems, optimizing routes, writing merchant tools. They built what they needed.
But this isn't a general-purpose foundation model that threatens GPT-5.5 or Claude Opus across all tasks. It's a proof of concept with a very sharp point: China can train competitive AI without American chips.
The Verification Problem
There's a catch. LongCat-2.0 is on HuggingFace, but the model weights aren't fully available for independent testing. The benchmarks come from Meituan's own evaluation. The AI community is treating the claims with cautious optimism — the SWE-bench scores are impressive if real, but without independent verification, skepticism is warranted.
This is a recurring pattern with Chinese AI announcements. DeepSeek made waves in early 2025 with claims of training efficiency that later held up under scrutiny. But other Chinese labs have announced "breakthroughs" that evaporated when independent researchers tried to replicate them.
🔥 Hot Takes
1. The US chip blockade is officially obsolete. Washington's entire strategy assumed that cutting off Nvidia would freeze Chinese AI development for years. LongCat-2.0 proves that assumption was wrong. Chinese chipmakers have reached the threshold where they can train trillion-parameter models at competitive quality. The gap hasn't closed completely — Nvidia still dominates training efficiency — but it's closed enough that export controls are now a speed bump, not a wall.
2. Meituan is the most interesting AI company nobody talks about. While the world obsesses over OpenAI, Anthropic, and DeepSeek, a Chinese food delivery company built a trillion-parameter model in two years. Meituan has the data (billions of delivery transactions, merchant interactions, user behavior), the compute (via partnerships with Chinese cloud providers), and the incentive (automating their massive logistics network). Don't be surprised if Meituan's AI division gets spun out at a $50B valuation within 18 months.
3. The real winner here isn't China — it's ASICs over GPUs. LongCat-2.0 was trained on "domestically made AI ASICs," not GPUs. This validates what Google figured out years ago with TPUs and what Amazon, Microsoft, and Meta are all racing to build: specialized AI chips are more efficient than general-purpose GPUs for training. Nvidia's moat isn't just manufacturing — it's CUDA, the software ecosystem that locks developers into their hardware. But if Chinese companies can build competitive models on non-CUDA ASICs, Nvidia's software advantage starts to look less impregnable.
What Happens Next
Expect three things:
First, more Chinese companies will announce domestic-chip training runs. ByteDance, Alibaba, Baidu, and Tencent all have the motivation and the resources. LongCat-2.0 just proved it's possible — now everyone will try.
Second, Washington will face pressure to escalate. If export controls aren't working, the policy options get more aggressive: sanctions on Chinese chipmakers, restrictions on AI model weights, or even attempts to limit cloud computing access. But each escalation risks pushing more countries toward Chinese AI infrastructure.
Third, the AI training cost curve is about to get weird. If Chinese ASICs can train trillion-parameter models at competitive quality, the global cost of training frontier models could drop dramatically. That benefits everyone — except maybe Nvidia's margins.
The bottom line: LongCat-2.0 isn't the best AI model in the world. But it's the most geopolitically significant model of 2026. It proves that America's chip blockade has failed, that Chinese AI can compete without Silicon Valley's hardware, and that the next phase of the AI race will be fought on hardware that Washington can't control.