Chinese food delivery and e-commerce company Meituan has published the full weights of LongCat-2.5 on Hugging Face under the MIT licence, letting developers run, modify and build on the model commercially. The weights were released on 9 November 2026, one week after Meituan introduced the model on 2 November.

LongCat-2.5 is a native multimodal mixture-of-experts model built for long-horizon agent tasks, work that stretches across many steps, such as developing a codebase, building a full project or creating 3D designs. Meituan says it has 1.6 trillion total parameters, with about 48 billion activated for each token, and a context window of one million tokens.

Benchmark scores close to closed frontier models

Meituan reports a sharp jump over its predecessor. On Terminal-Bench v2.1, which tests an AI agent's ability to complete real command-line tasks, LongCat-2.5 scores 88.4 against 70.8 for LongCat-2.0, and on DeepSWE v1.1 it scores 64.6 against 9.7. That puts it within about one point of GPT-5.6-Sol and Claude Opus 5 on Terminal-Bench v2.1, according to the company's own figures.

Those scores are self-reported by Meituan and have not been independently validated, so they should be treated as the company's claims rather than confirmed results. One discrepancy is already flagged: Hugging Face repository metadata lists the model at 1.8 trillion parameters, against Meituan's stated 1.6 trillion, which makes hardware sizing uncertain for teams planning to self-host.

Open weights, but a very large model to run

The MIT licence permits commercial use, modification and redistribution, subject to its notice requirements. Meituan also offers hosted access, with limited-time rates of 0.30 US dollars per million uncached input tokens and 1.20 dollars per million output tokens.

Self-hosting remains a heavy undertaking. The checkpoint is about 1.8 terabytes, and the full weight set must be available even though only a fraction of the parameters activate per token. Meituan points operators to an SGLang-FluentLLM deployment guide for neural processing units, with four-bit W4A4 quantisation aimed at Chinese accelerators. The release adds to a run of large open-weight agent models from Chinese companies competing with DeepSeek, Kimi and GLM.

Source: AI Weekly