Meituan has released LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese hardware and never processed on an Nvidia GPU.
The weights for the model are now public. The LongCat team at the Chinese food delivery firm announced the release one week after the initial unveiling. The system is designed for agents that handle long, multi-step tasks autonomously, with a specific focus on coding.
Meituan states the model performs at a level comparable to US counterparts like Claude Opus 5 on recent coding benchmarks.
Technical specifications and architecture
LongCat-2.5 uses a mixture-of-experts architecture. This means only a subset of the model’s specialised subnetworks activates for each token processed. While the total parameter count is 1.6 trillion, roughly 48 billion are active per token.
Through the API, the system handles up to one million tokens at once and can generate responses up to 128,000 tokens long. It accepts text, images, and video as input, responds in text, and allows users to toggle reasoning features on and off.
The model builds on its predecessor, LongCat-2.0, by adding image understanding. Many multimodal models attach a separate image encoder afterwards. LongCat-2.5 instead converts images into tokens that the model processes alongside text, following the standard practice for current large language models. This approach draws on the research model LongCat-Next and aims to improve visual coding, design, and 3D tasks.
Training methodology
For reinforcement learning, the team used real software tasks from GitHub pull requests along with recreated agent failures. To ensure the model actually solved the tasks during training, developers blocked internet access in the training environments, removed Git history and cached files, and had a language model check the solutions for tricks.
The team also distilled the capabilities of several specialised teacher models for agent work, reasoning, and dialog into LongCat-2.5.
Hardware requirements and performance gains
Meituan quantises the experts’ weights and activations to 4-bit precision. Chinese chips can process this format natively. In load tests, throughput rose by 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version. Meituan says accuracy stays on par with the uncompressed model.
A sample deployment runs on 194 Huawei Ascend chips from the Atlas A2 series. Two handle image processing, 64 handle prompt ingestion, and 128 handle response generation. There is no guidance for Nvidia GPUs.
Benchmark results
The LongCat team reports 88.4 points on Terminal Bench v2.1, a benchmark that measures how well agents solve command-line tasks. Meituan lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash at the top of the table with 90.6.
DeepSWE tests how well models fix bugs in real code repositories. Here LongCat-2.5 jumps from 9.7 to 64.6 points, placing it behind every other model in the table except Claude Opus 4.8, with Deepseek-V4.1-Flash leading at 74.2 and GPT-5.6 Sol at 72.7. The more recent Claude Opus 5.5 and GPT-6.1 Sol are not included in the table.
Terminal Bench v4.0, the harder variant, puts LongCat-2.5 at 38.5 points, slightly above GPT-5.6 Sol’s 37.3 but well below Opus 5’s 53.9. The model posts 91.9 on the web search benchmark BrowseComp, the highest score in the table, though Meituan ran this test in its own environment.
Availability and pricing
The weights are available under the MIT license, which allows commercial use, modification, and redistribution. Users can also try LongCat-2.5 in a chat interface at https://longcat.ai/. Its API supports both the OpenAI and Anthropic formats and plugs into agent harnesses like Claude Code, Codex, and Hermes.
At launch, Meituan is charging a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens.
Meituan says it trained the predecessor LongCat-2.0 in late June entirely on a cluster of more than 50,000 Chinese AI chips. Before that, the model ran anonymously as “Owl Alpha” on OpenRouter. Chinese open-weight models like Kimi K3.1 and GLM-5.3 have largely caught up with top US models on common benchmarks.
Open-weight models optimised for Chinese chips from the start make that hardware more attractive outside China too. Beyond multimodal agent models, Meituan is also working on image generators. The company previously released LongCat-Image, an open-source text-to-image model that gets by with just 6 billion parameters and prioritises efficiency over sheer model size.
