Meituan’s LongCat-2.0 Was Trained Entirely on Chinese Chips, the Company Says

6 min read · Last updated 2026-06-30

The most striking thing about Meituan’s new AI model is not its size, though it is enormous, but the hardware it ran on. The Chinese delivery and services giant launched LongCat-2.0 this week and says it is the first model of its scale trained entirely on domestically developed chips, a claim aimed squarely at the US export controls that have kept the best American silicon out of Chinese hands.

Why It Matters

For years the open question over China’s AI sector has been whether it can build frontier-scale models without Nvidia. Washington restricts exports of the most advanced chips on national security grounds, betting that limited access to cutting-edge silicon would slow China’s progress. A 1.6-trillion-parameter model that Meituan says was both trained and served on home-grown hardware is a direct test of that bet. If the claim holds, the single biggest lever the US has used to contain Chinese AI looks less decisive than it did.

What’s New

LongCat-2.0 carries 1.6 trillion parameters and a context window of one million tokens, and Meituan says its performance is comparable to Google’s Gemini 3.1 Pro, released in February. The company describes it as “the industry’s first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster.” The model has been open-sourced, putting the weights in the hands of anyone who wants to run or scrutinise them.

The crucial phrase is “end-to-end.” Plenty of Chinese models already run inference, the comparatively light task of answering a query once a model is built, on domestic hardware. Pre-training is the heavy part, the computationally brutal process in which a model digests vast data sets to learn its basic patterns, and it is where the most advanced chips have mattered most. Meituan’s claim that LongCat-2.0 was both pre-trained and served on domestic silicon is what makes the announcement more than a marketing line.

The Numbers

• 1.6 trillion parameters, putting LongCat-2.0 among the largest models publicly announced.
• 1 million token context window, for long-document and long-session work.
• 50,000-chip domestic compute cluster, used for what Meituan calls end-to-end training and inference.
• Comparable to Google Gemini 3.1 Pro, by Meituan’s own account, on the benchmarks it cites.
• Fully open-sourced weights, available for anyone to run or scrutinise.

“The industry’s first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster.”

What Comes Next

Independent verification will come from the open-source community, which can now run LongCat-2.0 against the benchmarks Meituan cites and test whether it genuinely matches a model like Gemini 3.1 Pro. The training-hardware claim is harder for outsiders to confirm, since it rests on Meituan’s account of its own infrastructure, and that caveat is worth holding in mind alongside the company’s confidence. LongCat-2.0 is the software counterpart to a broader hardware push: China recently claimed the supercomputing crown without US chips, and domestic challengers such as Alibaba’s T-Head unit are promoting home-grown accelerators like the Zhenwu M890 GPU.

What This Means in Practice

For anyone building on AI, the story is a reminder that the supply of capable models is globalising, not narrowing. An open-source model at this scale lowers the cost of frontier capability and widens the pool of providers beyond the familiar US names. Meituan itself is an unlikely flag-bearer, better known for food delivery than frontier AI, and its motive is concrete: routing, demand forecasting, and customer service all run on compute, and a model trained on domestic silicon insulates that compute from the next turn of the export-control screw. The practical takeaway for teams elsewhere is to keep an eye on open-weight models from outside the US, because the best price-to-performance option may increasingly come from an unexpected source.

The Bigger Picture

At its base, the AI contest between China and the United States has become a race over chips. Export controls were designed to widen America’s lead by denying China the hardware to train the largest models. Every credible claim of a frontier-scale model trained on domestic silicon chips away at that strategy. Meituan’s announcement is one more data point in a contest Washington built its restrictions to win, and that Beijing is determined to prove it can run on its own terms.

Frequently Asked Questions