Meituan releases LongCat-2.5, a 1.6-trillion-parameter AI model trained entirely on Chinese compute

Chinese local-services and delivery giant Meituan on November 2 released LongCat-2.5, its new-generation large model. According to an Economic Observer report republished in Meituan's newsroom, the model is a native multimodal system built for long-horizon agent tasks, with 1.6 trillion total parameters, and its training was carried out entirely on domestic Chinese computing hardware.

"Domestic models on domestic chips"

In June, Meituan said its LongCat-2.0 was the first trillion-parameter model to be trained and served end to end on Chinese chips. The new version strengthens native visual understanding and agentic coding, and improves the efficiency of long-context training and inference on domestic hardware.

Meituan said mid-stage training used sequences of up to 1 million tokens. To work around the smaller memory of individual domestic accelerators, the company worked with chip vendors to optimise memory use and inter-card communication, reporting an 86% gain in per-card training throughput over a standard distributed-data-parallel baseline on the long-context fine-tuning tasks it tested. For inference, it quantised both weights and activations to 4 bits, which it said raised throughput by about 40% compared with an 8-bit baseline.

Performance and use

On Meituan's own published evaluations, LongCat-2.5 scored between Anthropic's Claude Opus 4.8 and Claude Opus 5 on several mainstream benchmarks and beat Claude Opus 5 on some. These are company-reported results that have not yet been independently verified.

The LongCat API platform's change log says the model combines visual understanding with agentic coding, is designed to handle complex codebases and full-project builds through multi-step reasoning and tool use, and works with agent harnesses including Claude Code, Codex and Hermes.

Meituan CEO Wang Xing said in August that the company "will not become a token factory" and that its models will support its core business. Meituan has begun offering merchants an AI business workbench built on LongCat models.

Sources: Meituan newsroom (Economic Observer), LongCat API change log, LongCat blog