Meituan has published the weights of LongCat-2.5, a 1.6-trillion-parameter model designed for AI agents that work autonomously on long, multi-step tasks, especially programming. The release follows the official unveiling by one week and comes with a detail that matters more than any score: the model was trained entirely on Chinese hardware, without Nvidia GPUs.

Psst!Add AlexTech.ai asPreferred Source on Google

LongCat-2.5 uses a Mixture-of-Experts architecture, in which only part of its specialised subnetworks is active at any moment. Of its 1.6 trillion total parameters, about 48 billion are engaged for each token processed. The model understands text, images and video, answers in text, and lets users switch extended reasoning on or off. Compared with its predecessor LongCat-2.0, the main addition is built-in visual understanding: rather than attaching a separate image encoder afterwards, the model turns images into tokens processed alongside text, in line with current practice for large language models. The stated aim is better work on visual code, design and three-dimensional tasks.

official image of LongCat-2.5

Image: tech.meituan.com — https://tech.meituan.com

On results, Meituan reports 88.4 points on Terminal Bench v2.1, a test of how agents handle command-line tasks. That is slightly below the 89.1 attributed to Claude Opus 5 and the 88.8 for GPT-5.6 Sol, and well above the predecessor's 70.8. The sharpest jump concerns DeepSWE, which tests bug fixing in real code repositories: here the model rises from 9.7 to 64.6, still behind DeepSeek-V4.1-Flash (74.2) and GPT-5.6 Sol (72.7). On the harder Terminal Bench v4.0 it scores 38.5, against 53.9 for Opus 5. Its 91.9 on the BrowseComp web-search test is the highest in the table, but Meituan ran that evaluation in its own environment.

The technical core of the story is training. For reinforcement learning, the team used real software tasks drawn from GitHub pull requests, along with recreated agent failures. To stop the model from finding shortcuts, the developers cut internet access in the training environments, removed Git history and cached files, and had a separate language model check the solutions. The capabilities of several specialised teacher models, for agent work, reasoning and dialogue, were then transferred into LongCat-2.5 through distillation.

The most consequential part for anyone deploying it concerns hardware. Meituan quantised the expert weights and activations to 4-bit precision, a format Chinese chips can process natively. In its load tests, prompt processing throughput rises 43.3% and response generation 42.77% compared with an 8-bit version, with accuracy said to match the uncompressed model. The setup guide's sample deployment uses a cluster of 194 Huawei Atlas A2 accelerators: two for image processing, 64 for prompt ingestion and 128 for response generation. There is no guidance for Nvidia GPUs.

official image of LongCat-2.5

Image: tech.meituan.com — https://tech.meituan.com

The model is available under the MIT license, which permits commercial use, modification and redistribution, and can be tried in a chat interface. Its API supports both the OpenAI and Anthropic formats and works with agent tools such as Claude Code, Codex and Hermes. For the launch, Meituan is charging a limited-time promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens. The predecessor LongCat-2.0 was trained in late June and had earlier circulated on OpenRouter under the name "Owl Alpha".

The release reflects a broader shift: Chinese open-weight models have largely caught up with leading US models on common benchmarks, and some are now optimised for domestic chips from the outset. Meituan is also working on image generators, aiming at a model of just 6 billion parameters that puts efficiency ahead of sheer size.