What changed

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model designed for AI agents. The model was trained entirely on Chinese hardware and performs at the level of US models like Claude Opus 5 on coding benchmarks. It is a mixture-of-experts model that activates only a subset of its specialized subnetworks for each token it processes, with roughly 48 billion active parameters per token.

What this means in practice

LongCat-2.5 is built for agents that work through long, multi-step tasks autonomously, particularly coding. It understands text, images, and video, responds in text, and allows users to toggle reasoning on and off. The model converts images into tokens that are processed alongside text, as is standard for today's large language models. It also handles up to one million tokens at once and generates responses up to 128,000 tokens long.

The model is available under the MIT license, allowing commercial use, modification, and redistribution. Users can try LongCat-2.5 in a chat interface at https://longcat.ai/ or use its API, which supports OpenAI and Anthropic formats. Meituan is charging a limited-time promotional price of approximately €0.27 per million input tokens and €1.07 per million output tokens.

Limitations and availability

LongCat-2.5 is available in a separate repository and requires a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series for self-hosting. Two chips handle image processing, 64 handle prompt ingestion, and 128 handle response generation. There is no guidance for Nvidia GPUs. The model trails Claude Opus 5 slightly on terminal tasks, scoring 88.4 on Terminal Bench v2.1, compared to Opus 5's 89.1. It also performs well on the web search benchmark BrowseComp, achieving the highest score in the table.

The supplied announcement does not specify the exact conditions under which the model was tested or the exact hardware configurations used beyond the Huawei cluster. The model's performance on various benchmarks is reported by Meituan, but independent testing is not mentioned.

Source facts for Meituan releases 1.6-trillion-parameter agent model trained on Chinese hardware: It was trained entirely on Chinese hardware and performs at the level US models like Claude Opus 5 reached a few months ago on coding benchmarks. The LongCat team at Chinese food delivery company Meituan has released the weights for LongCat-2.5, one week after the model's official unveiling. LongCat-2.5 is built for agents that work through long, multi-step tasks autonomously, particularly coding.
Original AINewsTrend explainer based on the cited source. This graphic summarizes source statements; it is not an independent benchmark.

Sources and further reading

  1. thedecoder: Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU