On June 30, 2026, Chinese internet company Meituan officially released and open-sourced the next-generation trillion-parameter model LongCat-2.0. As the industry's first trillion-parameter model to complete end-to-end training and inference on a ~50,000-card purely domestic compute cluster, LongCat-2.0 has a total parameter scale of 1.6T (average active ~48B, dynamic range 33B~56B) and natively supports a 1M ultra-long context. This major release not only marks a technical leap for Chinese companies in ultra-large-scale model training but also provides an important practical case for autonomous and controllable domestic AI infrastructure.

LongCat-2.0 adopts an efficient mixture-of-experts (MoE) architecture and is deeply optimized for agentic encoding, complex task planning, and multi-modal scenarios. Its open-source release further lowers the barrier for the industry to access cutting-edge models and is expected to accelerate iteration and innovation in the domestic AI application ecosystem.

Technical highlights: a balanced breakthrough in scale, efficiency and scenario fit

The core innovations of LongCat-2.0 are reflected in the following aspects:

Massive-scale MoE architecture: Total parameters reach 1.6T, with about 48B parameters dynamically activated during inference (range 33B~56B). This sparse-activation design preserves large capacity while significantly reducing per-inference compute, achieving a favorable balance of performance and efficiency.
Native 1M ultra-long context support: Single inferences can handle input on the order of millions of characters, far exceeding most international mainstream models. This capability is especially suited for long-document analysis, codebase understanding, and multi-turn complex dialogue scenarios.
Agentic capability optimization: The model has targeted enhancements for agent applications, demonstrating strong performance in code generation, task planning, and tool invocation, making it suitable for enterprise automation and complex workflows.
Full-stack adaptation to domestic compute: Training and inference were conducted entirely on domestic compute clusters (around 50,000–60,000 cards). This is currently the largest reported case of a domestic compute-trained trillion-parameter model, validating the usability of the local hardware ecosystem for ultra-large-scale training.

The Meituan LongCat team says the model achieved competitive results on benchmarks such as SWE-bench Pro and Terminal-Bench, and demonstrated high deployment efficiency in real enterprise scenarios.

Context of release: breakthroughs in domestic compute and industrial autonomy needs

There is a deep industrial background to the birth of LongCat-2.0:

Urgency of domestic compute autonomy: Against increasing uncertainty in the international chip supply chain, Meituan chose to train the trillion-parameter model entirely using domestic compute, which carries important demonstrative significance. Although domestic large models have been released before, full-process training of a trillion-parameter model on domestic compute is a first at this scale.
Meituan's business-driven needs: As a local services giant, Meituan has rich experience in AI application scenarios. The development of LongCat-2.0 directly serves its agentic application needs—such as intelligent customer service, content generation, and supply-chain optimization—driving intelligent upgrades across its business.
Open-source strategic choice: Unlike some international giants that keep models closed, Meituan chose to open-source LongCat-2.0 to pool community efforts, accelerate the construction of China’s AI ecosystem, and broaden its technica