Meituan releases LongCat-2.0: an open 1.6-trillion-parameter MoE model with a 1-million-token context

Chinese technology company Meituan released LongCat-2.0 on July 5, 2026: an open Mixture-of-Experts model with 1.6 trillion parameters and a native context window of 1 million tokens, trained and deployed entirely on domestic AI accelerators.

What's inside LongCat-2.0

LongCat-2.0 is built on a Mixture-of-Experts architecture: for each token it processes, only about 48 billion of the 1.6 trillion parameters are activated. This approach lets the model keep the quality of dense models of comparable scale while cutting inference compute substantially. It calls on the relevant expert blocks selectively instead of "spending" the entire compute budget on every token.

Key characteristics of the model:

• Total parameters: 1.6 trillion; active per token: ~48 billion
• Native context: 1 million tokens
• Attention mechanism: LongCat Sparse Attention (Meituan's own development)
• Infrastructure: training and inference on domestic ASIC-based AI supercomputing pods
• Status: open model, available via API

Why a 1-million-token context changes the rules of the game

A context window of 1 million tokens is among the longest of any open model today. In practice, this means being able to process entire code repositories, large legal or financial documents and long conversation histories in a single request, without splitting them into chunks and losing coherence.

To keep a context this long computationally manageable, Meituan developed its own mechanism, LongCat Sparse Attention. It reduces the quadratic complexity of standard self-attention: instead of full pairwise attention among all tokens, the model uses sparse patterns that cut the amount of computation without a significant loss of quality on long sequences.

Sovereign infrastructure: the full cycle without foreign GPUs

One notable detail of the release is that Meituan ran the entire cycle, from training to production inference, on supercomputing pods built with domestic ASIC-based AI accelerators. Against the backdrop of US export restrictions on high-performance GPUs, this shows that large Chinese companies are not merely adapting to infrastructure barriers but are building competitive frontier-class products on their own compute base.

Meituan is known primarily as a delivery and lifestyle-services platform rather than a traditional AI lab. That makes it all the more significant that the company closed the loop, from silicon to an open model with 1.6 trillion parameters, without using foreign hardware. Until now, public reports of full-scale training of models at this size without NVIDIA H100s have been rare among non-specialized technology companies.

What this means

The release of LongCat-2.0 expands the pool of open MoE models with extremely long context. It also confirms a trend: large Chinese technology companies outside the classic AI research circle have learned to produce competitive frontier models, and they are doing it on their own hardware. For developers, it adds one more open option with a 1-million-token context for tasks where sequence length is critical.

Frequently asked questions

How many parameters does

LongCat-2.0 actually use at runtime?

For each token it processes, about 48 billion of the 1.6 trillion parameters are activated. This is the standard MoE principle, which keeps quality high while holding inference compute costs manageable.

How did Meituan achieve a 1-million-token context?

The company developed its own mechanism, LongCat Sparse Attention, which reduces the quadratic complexity of standard attention by using sparse patterns to process long sequences.