Clarity for AI buildersSponsored by AI Routing
availability

Meituan releases weights for LongCat-2.5, a 1.6-trillion-parameter model trained on Chinese chips

Meituan releases weights for LongCat-2.5, a 1.6-trillion-parameter model trained on Chinese chips

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese hardware.

The Decoder reports that the weights are available under the MIT license, which permits commercial use, modification and redistribution. LongCat-2.5 is designed for multi-step tasks such as fixing code in a repository. It uses a mixture-of-experts architecture: although the model has 1.6 trillion parameters in total, it activates roughly 48 billion per token.

Through Meituan’s API, it can process up to one million tokens at once and generate up to 128,000 tokens in a response. It accepts text, images and video and replies in text.

Meituan’s benchmark results show that how closely LongCat-2.5 approaches competing models depends on the task. On Terminal Bench v2.1, it scored 88.4, just behind Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. On the harder Terminal Bench v4.0, its 38.5 put it slightly ahead of GPT-5.6 Sol at 37.3 but well behind Claude Opus 5 at 53.9. On DeepSWE, which tests bug fixes in real repositories, LongCat-2.5 improved from its predecessor’s 9.7 to 64.6, but trailed Deepseek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7.

The comparison did not include the newer Claude Opus 5.5 or GPT-6.1 Sol.

According to The Decoder, Meituan used tasks drawn from GitHub pull requests and recreated agent failures during reinforcement learning. The team blocked internet access and removed Git history and cached files from the training environments, limiting ways the model could retrieve a solution rather than work through the task. The model also processes images as discrete tokens alongside text, an approach Meituan says is intended to help with tasks such as visual coding and design.

The permissive license allows self-hosting, but Meituan’s documented example uses a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. It assigns two accelerators to image processing, 64 to prompt processing and 128 to response generation. The guide does not cover Nvidia GPUs. Meituan also released a separate 4-bit version and reported that, in its load tests, throughput rose 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version.

Meituan says the compressed model maintained the accuracy of the uncompressed model.