Meituan releases weights for 1.6-trillion-parameter LongCat-2.5 model built for agents
China’s Meituan has released the weights of its LongCat-2.5 model under the MIT License. In Meituan’s comparison, the model comes close to Claude Opus 5 on some coding and agent benchmarks but falls clearly behind Opus 5 on the Terminal-Bench 4.0 command-line benchmark.
On 15 November, The Decoder reported that the LongCat team at Meituan, a Chinese food delivery company, had released the model weights a week after the official launch. LongCat-2.5 is built for AI agents that carry out long, multi-step tasks on their own, especially programming.
LongCat-2.5 is a mixture-of-experts (MoE) model: it uses only some of its specialised sub-networks to process each token. Meituan’s model card says about 48 billion of the model’s 1.6 trillion parameters are active per token. The model understands text, images and video and responds in text. It splits images at native resolution into separate tokens and processes them alongside text.
The Decoder reported that the model was trained entirely on Chinese hardware. In the model card, Meituan says it optimised the infrastructure for domestic AI accelerators.
For reinforcement learning, Meituan used programming tasks compiled from GitHub pull requests, The Decoder wrote. To stop the model taking shortcuts, Meituan cut off internet access to the training environments and removed the Git history. A language model checked the solutions for tricks.
Close to the top in Meituan’s comparison, behind Opus 5 on Terminal-Bench 4.0
In Meituan’s comparison, LongCat-2.5 clearly improves on its predecessor and comes close to Claude Opus 5 and GPT-5.6 Sol on some agent benchmarks. The LongCat figures are Meituan’s own measurements, and the company says they are only indicative.
The table shows scores on four benchmarks. Terminal-Bench measures command-line tasks, DeepSWE bug fixing in real repositories and BrowseComp web search.
| Benchmark | LongCat-2.5 | LongCat-2.0 | Claude Opus 5 | GPT-5.6 Sol | DeepSeek-V4.1-Flash |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.4 | 70.8 | 89.1 | 88.8 | 90.6 |
| Terminal-Bench 4.0 | 38.5 | 3.0 | 53.9 | 37.3 | 31.2 |
| DeepSWE 1.1 | 64.6 | 9.7 | 68.8 | 72.7 | 74.2 |
| BrowseComp | 91.9 | 79.9 | 90.8 | 90.4 | 91.5 |
Source: Meituan’s model card, in points. Meituan measured the LongCat figures; competitor figures come mainly from public reports. Meituan measured DeepSeek’s BrowseComp result.
The table does not show how LongCat-2.5 compares with the latest models: The Decoder noted that Claude Opus 5.5 and GPT-6.1 Sol are missing from the comparison. On BrowseComp, LongCat-2.5 achieved the highest score in the table. However, Meituan ran the benchmark with its own harness, the program that controls the agent, gives the model its tools and runs its work steps.
4-bit version and deployment
A separate repository holds a version whose expert weights and activations are quantised to 4 bits. Chinese accelerators compute with these numbers directly. Meituan says accuracy on the evaluated benchmarks roughly matches the unquantised model. In the company’s stress tests, throughput rose by 43.3 per cent in prefill and 42.77 per cent in decode compared with the 8-bit version.
The model card says the model can run on NPU AI accelerators and points to an SGLang-FluentLLM guide on GitHub. The Decoder reported that the guide’s example cluster has 194 Huawei Ascend Atlas A2 accelerators and that there are no instructions for Nvidia GPUs.
The MIT License allows commercial use and modification. The model can be tried in the longcat.ai chat service. The Decoder said the API supports OpenAI and Anthropic formats and works with tools such as Claude Code and Codex. The API has a context window of 1 million tokens and a limited-time promotional price of about €0.27 per million input tokens and €1.07 per million output tokens.
Read more: Models