Meituan Releases LongCat-2.0: Open MoE Model with 1.6 Trillion Parameters and 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: when processing each token, only about 48 billion of the 1.6 trillion parameters are activated. This approach preserves the quality of dense models at comparable scale while requiring substantially lower computational costs for inference — the model precisely engages the necessary expert blocks rather than "spending" the entire computational budget on each token.

Key model characteristics:

• Total parameter count — 1.6 trillion, active per token — ~48 billion
• Native context — 1 million tokens
• Attention mechanism — LongCat Sparse Attention (Meituan's proprietary development)
• Infrastructure — training and inference on domestic AI superpods based on ASIC
• Status — open model, available via API

Why a 1 Million Token Context Changes the Game

A context window of 1 million tokens is among the longest in open models to date. In practice, this enables processing entire code repositories, voluminous legal or financial documents, lengthy conversation histories — within a single request, without chunking and without losing coherence.

To keep such long context computationally manageable, Meituan developed its own LongCat Sparse Attention mechanism. It reduces the quadratic complexity of standard self-attention: instead of full mutual attention among all tokens, the model applies sparse patterns that reduce computational volume without significant loss of quality in processing long sequences.

Sovereign

Infrastructure: a Complete Cycle Without Foreign GPUs

One notable detail of the release is that Meituan conducted the entire cycle from training to production inference on superpods with domestic ASIC-based AI accelerators. Against the backdrop of American export restrictions on high-performance GPUs, this demonstrates: major Chinese technology companies are not simply adapting to infrastructure barriers, but creating competitive frontier-class products on their own computational foundation.

Meituan is primarily known as a delivery and lifestyle services platform, not as a traditional AI laboratory. All the more significant, then, that the company closed the cycle — from silicon to an open model with 1.6 trillion parameters — without relying on foreign equipment. Previously, public reports of full-scale training of models of such magnitude without NVIDIA H100s were 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 and simultaneously solidifies a trend: major Chinese technology companies outside the traditional AI research establishment have learned to produce competitive frontier models — and they're doing it on their own hardware. For developers, there's another open option with a 1 million token context for tasks where sequence length is critical.

Frequently Asked Questions

How many parameters are actually used when LongCat-2.0 runs?

When processing each token, about 48 billion parameters out of 1.6 trillion are activated — this is the standard principle of MoE architecture, allowing high quality while maintaining manageable computational costs for inference.

How did Meituan achieve a 1 million token context?

The company developed its own LongCat Sparse Attention mechanism, which reduces the quadratic complexity of standard attention through sparse patterns for processing long sequences.