Meituan launches LongCat-2.5 multimodal model for AI agents

LongCat-2.5 is Meituan’s new native multimodal model for long-horizon agent work, with the company pitching it as a system that can handle visual understanding, software engineering and extended context on domestic Chinese compute. The announcement gives the model 1.6 trillion total parameters and says developers can access it through an API, while users can try it through longcat.ai. ¹

What Meituan is claiming

Meituan says LongCat-2.5 improves on LongCat-2.0 in native visual understanding, agentic coding and long-context training and inference. The model can reportedly process visual input as tokens in a shared representation, which is intended to help with tasks such as interactive web pages and 3D presentations. The company also says large-scale long-horizon reinforcement learning has been used to improve end-to-end engineering work, rather than simply producing isolated snippets of code. ¹

The announcement says the model was trained entirely on Chinese domestic compute. Meituan describes a training setup that can use sequences of up to 1 million tokens during the mid-training stage, with a total training volume running into several trillion tokens. The company says it worked with domestic AI accelerator suppliers to optimise memory use, communication between cards and the balance between computation and data transfer. ¹

That work is aimed at a practical problem: long-context training is expensive, and smaller accelerator memory can make moving data between cards a bottleneck. Meituan says its super-node optimisations keep more of the parameter and task-data traffic inside a high-speed interconnect, reducing the amount of communication that has to travel further through the system. ¹

The performance figures

Meituan reports that an optimised single-card training setup delivered 86 per cent higher throughput than a DDP, or distributed data-parallel, baseline in the long-context supervised fine-tuning task it measured. In simple terms, the company says one card could complete about 1.86 times the work in the same period under that test. This is a company-reported result rather than an independent benchmark. ¹

For inference, LongCat-2.5 uses a W4A4 approach that quantises both expert weights and activations to four bits. Meituan says the system was tuned for the characteristics of domestic AI accelerators and maintained performance comparable with the non-quantised model on the downstream tasks it tested. ¹

The company further reports that, compared with a W8A8 INT8 baseline, input processing and token-generation throughput rose by about 40 per cent. It says first-output wait time fell by about 16 per cent, while the time needed to generate later tokens fell by about 20 per cent. Those numbers describe Meituan’s own test conditions, so they should be treated as indicative rather than a guarantee of performance on every API workload. ¹

Built for Meituan’s own work

The announcement places LongCat-2.5 inside Meituan’s wider plan to use AI in its local-services business. Meituan says the model is intended to support commercial analysis and software development, including tools for local merchants. Its example describes an AI system that can build a multi-channel profit statement, spot a rise in transaction value alongside a fall in channel profit, and suggest different resource allocations based on stock, repeat purchases and clearance value. ¹

Meituan also points to its AI operations workspace, Miaoshou, which is designed to help merchants understand business metrics and brand rules before calling tools and systems within the permissions they grant. The company says the workspace is already being used by several chain brands, although the announcement does not provide independently audited results for those deployments. ¹

Our opinion

LongCat-2.5 is interesting because Meituan is presenting it as infrastructure for real work rather than a chatbot launch in isolation. The combination of native multimodal input, long-context training and aggressive low-bit inference is aimed at the expensive parts of agentic software: seeing a complex interface, keeping track of a large task and producing output quickly enough to be useful.

The caveat is that the headline comparisons and benchmark positions remain Meituan’s claims. The official chart places LongCat-2.5 between different Claude and GPT systems across several tests, but the announcement does not give enough independent methodology to treat those comparisons as settled evidence. The more meaningful test will be whether developers can reproduce the reported efficiency gains through the API, and whether the model can complete long jobs reliably rather than merely score well on selected evaluations.

That makes the launch worth watching, especially for developers working with Chinese accelerator hardware or merchant-facing automation. It is not yet proof that LongCat-2.5 has overtaken the leading general-purpose models, but it is a clear statement of Meituan’s ambition: build a multimodal agent model around its own compute constraints and then put it to work inside the company’s core business.