Meituan drops 1.6-trillion-parameter LongCat-2.0 trained on Chinese chips

Meituan claims it trained the 1.6 trillion parameter model on domestic Chinese hardware, avoiding Nvidia GPUs altogether. The company is the biggest platform in China for local services and food delivery.

The release lands as U.S. export controls keep reshaping how Chinese firms build large-scale AI. Meituan trained LongCat-2.0 on domestic ASIC superpods. The company frames the model as proof Chinese firms can hit frontier scale without Nvidia’s CUDA-based chips.

LongCat-2.0 runs on a sparse mixture-of-experts design

LongCat-2.0 uses a sparse mixture-of-experts architecture. DeepSeek and Mistral’s Mixtral use that same broad approach. Instead of simultaneously firing all 1.6 trillion parameters, an internal router selects a subset of specialized sub-models for each token. Compared to a dense model of the same size, that design keeps inference costs down.

The model ships with a one million token context window. Both DeepSeek-R1-0528 and OpenAI’s GPT-OSS have a maximum token value of 128,000. In the published benchmarks, Meituan compared LongCat-2.0 to Google’s, OpenAI’s, and Anthropic’s closed-source models. So far, those assertions have not been validated by impartial third-party assessments.

Meituan developed LongCat-2.0 to serve as the primary reasoning engine for AI agents and coding tools. The company pointed to code comprehension, repository-wide edits, and automated task execution as target use cases.

Bernstein pegs Nvidia at 40% of China’s AI chip market

An estimate from the equity research firm Bernstein for 2025 placed Nvidia’s share of the Chinese artificial intelligence chip market at ~40%. Huawei has a similar percentage. Bernstein predicted that Huawei would gain ground this year, causing Nvidia’s share to fall by 8 percentage points.

As for domestic ASIC clusters, Meituan claims to have trained and optimized LongCat-2.0. This means that the model doesn’t need Nvidia’s software stack and can instead run on hardware that is already present in China. Instead of disjointed third-party configurations, “superpods” implies fully integrated enterprise-grade hardware.

Neither consumer devices nor the majority of on-premises 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 model parallelism.

Delivery of meals 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 made its entry into the AI space. According to SiliconANGLE, it did not publicly announce its plans for internal model development until 2025.

MiniMax, another Chinese AI startup, drew backing from Alibaba and miHoYo. According to earlier reports from Cryptopolitan, these investors committed not to sell shares before the lock-up expiration on July 9.

MiniMax rolled out its own million-token-context model, M3, in early June 2026, per Cai Lian She as cited by Cryptopolitan. MiniMax has introduced prices that are significantly lower than the market leaders in the United States.

Independent testing will decide how seriously developers outside China take LongCat-2.0. Optimization for domestic chips may limit performance on Nvidia hardware, which still dominates data centers worldwide. The core reasoning architecture, according to Meituan, remains portable.

If you're reading this, you’re already ahead. Stay there with our newsletter.