Meituan Just Dropped a Trillion-Parameter Model Trained on 60,000 Domestic GPUs — Not a Single NVIDIA Chip in Sight
On April 24, Meituan quietly did something that should be on every AI developer’s radar.
No press conference. No hype cycle. Just a testing invite page for LongCat-2.0-Preview — their next-generation foundation model.
The specs: 1 trillion+ parameters. MoE architecture. 1 million token context window. Trained on a cluster of 50,000 to 60,000 accelerator cards.
But here’s the part that actually matters.
Every single one of those cards was made in China. Zero NVIDIA. Zero imports.
This is reportedly the largest AI model training run ever completed entirely on domestic Chinese compute infrastructure.
Let that satisfying detail sit for a moment.
Wait — Meituan? The Food Delivery Company?
Yes. The same company you’d associate with ordering bubble tea and booking hotels in China.
But if your mental model of Meituan is still “local services app,” you’re several chapters behind. Starting in 2024, founder Wang Xing began making an unmistakable pivot. During an earnings call, he dropped a line that raised eyebrows: “We will continue to invest billions of dollars to ensure adequate computing power.”
Billions. With a B. In USD.
Then came the investment spree. Meituan poured capital into domestic chip companies: Moore Threads, Muxi, Unisoc, Aixin Yuanzhi — all Chinese GPU and semiconductor firms. They also backed AI model companies like Zhipu AI and Moonshot AI (the makers of Kimi).
Silicon to compute to models. Meituan was quietly building an end-to-end domestic AI supply chain.
560 Billion to 1 Trillion — In Just 7 Months
Let’s trace the timeline, because it’s genuinely impressive.
September 2025: Meituan open-sourced LongCat-Flash — a 560B-parameter MoE model with 27B active parameters. Pre-trained on 20 trillion tokens in 30 days. It went semi-viral for its inference speed. Most people’s first reaction: “Wait, Meituan makes foundation models?”
Then came a rapid-fire model family:
- LongCat-Flash-Lite — lightweight variant for edge deployment
- LongCat-Flash-Thinking — enhanced reasoning capabilities
- LongCat-Next (March 2026) — native multimodal, 68.5B parameters
And now, LongCat-2.0-Preview: nearly 2x the parameter count of LongCat-Flash, jumping from 560B to the trillion-parameter club.
Seven months. From open-source underdog to trillion-parameter contender.
The Real Story: China Just Proved Domestic Compute Can Train Frontier Models
Trillion-parameter models aren’t exactly rare in 2026. DeepSeek V4, GPT-5.5, Qwen 3.6 Plus — everyone’s operating at this scale now.
But here’s what makes LongCat-2.0-Preview different: it was trained entirely on Chinese-made hardware.
50,000 to 60,000 domestically produced accelerator cards. End to end — pre-training through inference — with zero reliance on imported chips.
You might be thinking: haven’t Chinese companies been using domestic GPUs for a while now? What’s the big deal?
Here’s the distinction. Previous efforts were largely at the “it technically works” stage. LongCat-2.0-Preview proves something fundamentally different: domestic compute doesn’t just work — it works at frontier scale. The gap between “functional” and “production-grade” has been crossed.
In the context of ongoing U.S. chip export controls targeting China’s AI sector, this isn’t just a technical milestone. It’s a geopolitical signal. The export restrictions were designed to slow China’s AI progress by choking off access to cutting-edge chips. This training run suggests the strategy may be losing its bite.
It Launched the Same Day as DeepSeek V4 — Coincidence?
Here’s a fun fact. LongCat-2.0-Preview and DeepSeek V4 both opened testing on the exact same day — April 24.
Two trillion-parameter MoE models. Same day. Both supporting million-token context windows. Both from Chinese companies.
In China’s AI ecosystem, this kind of “coincidence” typically means one thing: nobody wants to be second. The race has entered its final lap, and the front-runners are sprinting.
The Chinese foundation model landscape is rapidly consolidating. The era of “a hundred models competing” is over. What’s left is a small group of players with the compute, data, and engineering depth to operate at the frontier. DeepSeek is one. Meituan just made a loud case that it’s another.
Why Agent-First Architecture Matters Here
LongCat-2.0-Preview was optimized for three things: code generation, complex task planning, and enterprise automation.
Notice the common thread? These are all AI agent workloads.
A 1-million-token context window + trillion parameters + efficient MoE routing = a model architecturally built for long-chain, multi-step agentic workflows. This isn’t a chatbot. It’s an execution engine.
Wang Xing has been explicit about the endgame: transforming the Meituan app into an “AI-powered app.” Picture an AI assistant embedded in an app used by tens of millions daily — one that understands your full conversation history, calls multiple backend systems, and autonomously orchestrates complex tasks.
And Meituan’s moat? It processes tens of millions of real orders every single day. That’s not a synthetic benchmark. That’s the world’s most demanding AI agent testing ground.
So What Does This Actually Mean?
Three takeaways:
Scale: Meituan has entered the trillion-parameter tier, placing it alongside the world’s leading foundation model labs.
Sovereignty: The largest AI training run ever completed on fully domestic Chinese hardware. In a world of chip export controls, this changes the equation.
Application: This isn’t a model built to top leaderboards. It’s built to power agent workflows inside one of the world’s highest-traffic consumer platforms.
Put it all together, and the narrative writes itself: A Chinese “food delivery” company just trained a trillion-parameter flagship model on entirely domestic chips — and plans to deploy it inside an app that serves tens of millions of people daily.
If you’re tracking the global AI race, LongCat-2.0-Preview deserves five minutes of your attention.
Meituan Just Dropped a Trillion-Parameter Model Trained on 60,000 Domestic GPUs — Not a Single… was originally published in DataDrivenInvestor on Medium, where people are continuing the conversation by highlighting and responding to this story.
