Meituan Unveils LongCat-2.0: A 1.6 Trillion-Parameter Open MoE Model for Agentic Coding

Meituan, a prominent technology company, has released LongCat-2.0, a large-scale Mixture-of-Experts (MoE) language model with 1.6 trillion total parameters. This model is notable for activating approximately 48 billion parameters per token and is specifically designed to excel in agentic coding tasks, encompassing code understanding, generation, and execution within agent workflows. LongCat-2.0 introduces a native 1-million-token context window, built upon its proprietary LongCat Sparse Attention mechanism. A significant detail for the global AI landscape, including its development implications for India, is that both the training and serving of LongCat-2.0 were conducted entirely on domestic AI ASIC superpods.

Key features and architecture

LongCat-2.0 represents Meituan’s next-generation open model, following its 560B predecessor, LongCat-Flash, released in 2025. The core architectural design prioritizes reliable and efficient agentic coding. Meituan reports that pretraining involved over 35 trillion tokens and millions of accelerator-hours, with no reported rollbacks or irrecoverable loss spikes—a stability claim that underscores its performance on non-Nvidia hardware. The model integrates several techniques to optimize scaling costs, including a 6D parallelism scheme, a prefill-decode disaggregated architecture for serving, and the use of ‘super kernels’ and L2-cache weight prefetching to mitigate I/O latency.

Performance and benchmarks

Meituan positions LongCat-2.0 as a specialized tool for agentic coding. The company’s internal testing indicates strong performance in software engineering tasks. On SWE-bench Pro, LongCat-2.0 reportedly surpassed GPT-5.5 with a score of 58.6. Meituan also claims its overall performance is comparable to Google’s Gemini 3.1 Pro, particularly in software engineering contexts. However, the source notes that its performance on broader general-agent benchmarks like FORTE and BrowseComp suggests it trails leading frontier systems, and independent leaderboard confirmation is not yet available. The model is specifically tuned for agent-style software work rather than general conversational AI.

Key facts

Feature	Detail
Total Parameters	6 Trillion (Mixture-of-Experts)
Activated Parameters	~48 Billion per token
Context Window	1 Million tokens (native, with LongCat Sparse Attention)
Training/Serving	Domestic AI ASIC superpods
Primary Use Case	Agentic Coding (understanding, generation, execution)

API access and pricing

LongCat-2.0 is accessible through the LongCat API Platform, which provides both OpenAI-compatible and Anthropic-compatible endpoints. It is also available via OpenRouter and integrated into platforms such as Claude Code, OpenClaw, OpenCode, and Codex. While local self-hosting is not yet an option as weights remain pending, developers can interact with the model using its OpenAI-compatible endpoint, specifying `LongCat-2.0` as the model ID. The maximum output length supported is 131,072 tokens (128K). Pricing is reported at $0.75 per million input tokens and $2.95 per million output tokens, with a launch promotion offering reduced rates of $0.30 and $1.20, respectively, and cached context reads being free. These figures are based on third-party coverage and are subject to change.

Implications for AI development in India

The release of LongCat-2.0, particularly its focus on agentic coding and training on domestic hardware, has implications for India’s burgeoning AI ecosystem. As India pushes for self-reliance in technology and promotes initiatives like the IndiaAI Mission, models developed on indigenous infrastructure demonstrate a pathway for reducing dependency on foreign hardware and software ecosystems. For Indian startups and developers working on AI agents and software engineering tools, LongCat-2.0 could offer a powerful, specialized foundation model. Its API accessibility allows for integration into existing development workflows, potentially accelerating innovation in areas like automated code generation, debugging, and complex software task execution, aligning with India’s growing demand for skilled AI and software professionals.

Source: MarkTechPost, https://www.marktechpost.com/2026/07/05/meituan-releases-longcat-2-0-a-1-6t-parameter-open-moe-model-with-native-1m-context-and-longcat-sparse-attention/