Meituan claims to have trained the 1.6 trillion parameter model on local Chinese hardware, avoiding Nvidia GPUs entirely. The company is China’s largest platform for local services and food delivery.
The release comes as US export controls continue to reshape how Chinese companies build AI at scale. Meituan trained LongCat-2.0 on hyperlocal ASICs. The company is framing the model as evidence that Chinese companies can reach the frontier scale without using Nvidia’s CUDA-based chips.
LongCat-2.0 employs a sparse design from a mix of experts
LongCat-2.0 uses a sparse architecture of expert mix. DeepSeek and Mistral’s Mixtral use the same broad approach. Instead of releasing all 1.6 trillion parameters at once, the internal router selects a subset of specialized submodels for each token. Compared to a dense model of the same size, this design keeps inference costs low.
The template comes with a context window of 1 million tokens. Both DeepSeek-R1-0528 and OpenAI’s GPT-OSS have a maximum token value of 128,000. In published benchmarks, Meituan compared LongCat-2.0 to closed-source models from Google, OpenAI, and Anthropic. To date, these assertions have not been verified by unbiased third-party evaluations.
Meituan He developed LongCat-2.0 to serve as the underlying logic engine for AI agents and programming tools. The company cited code understanding, repository-level editing, and automated task execution as targeted use cases.
Bernstein pegs Nvidia with 40% of China’s AI chip market
An estimate from equity research firm Bernstein for 2025 puts Nvidia’s share of the Chinese AI chip market at roughly 40%. Huawei has a similar ratio. Bernstein expects Huawei to achieve gains this year, which will lead to a decline in Nvidia’s share by 8 percentage points.
As for local ASIC groups, Meituan claims to have trained and improved LongCat-2.0. This means that the model does not need the Nvidia software stack and can instead run on hardware already in China. Instead of disjointed third-party configurations, “superpods” include fully integrated, enterprise-grade hardware.
Neither consumer devices nor the majority of domestic systems will be able to handle LongCat-2.0’s 1.6 trillion parameters. It resides in data centers, distributed across high-density inference clusters that use a parallelism model.
Meal delivery is Meituan’s claim to fame, not its development of frontier AI. By purchasing AI startup Light Year Beyond for $281 million in 2023, the Beijing company has entered the AI space. According to SiliconANGLE, it has not publicly announced its plans to develop the in-house model until 2025.
MiniMax, another Chinese AI startup, has received backing from Alibaba and miHoYo. According to the above Reports From Cryptopolitan, these investors have committed not to sell shares before the lockup period expires on July 9.
MiniMax rolled out its million-token context model, M3, in early June 2026, according to Cai Lian She as cited by Cryptopolitan. MiniMax offered prices well below those of market leaders in the United States.
Independent testing will determine how seriously developers outside China take LongCat-2.0. Optimizing local chips may limit the performance of Nvidia hardware, which still dominates data centers around the world. According to Mitwan, the basic structure of inference remains transferable.
Don’t just read cryptocurrency news. Understand that. Subscribe to our newsletter. It’s free.
