Behind LongCat-2.5’s performance, the rise of Chinese models

With 1.6 trillion parameters, LongCat-2.5 focuses on extended reasoning, visual understanding, and autonomous software development.

Chinese group Meituan is continuing the development of its LongCat family of models with a new version focused on carrying out complex tasks. Called LongCat-2.5, this native multimodal model combines visual understanding, programming, and agentic capabilities to carry out projects requiring many steps, sometimes without continuous human intervention. It is an approach that goes beyond simple response generation to focus on completing full tasks.

According to the technical presentation published by Meituan, LongCat-2.5 is based on a mixture-of-experts (MoE) architecture with 1.6 trillion parameters, of which about 48 billion are activated for each token processed. This setup makes it possible to mobilize only a fraction of the model for each operation rather than all of its capabilities at once. From software development to visual understanding

LongCat-2.5 was designed to handle long-duration agentic tasks, including development within large codebases, building complete applications, and the successive use of multiple tools. Meituan says it strengthened its training using longer execution trajectories and reinforcement learning, notably reusing failures encountered by its agents to create new verifiable exercises.

The other major development concerns visual information processing. The model represents images as multi-level discrete tokens, making it possible to process visual and textual content within a shared framework. This approach is aimed in particular at visual coding applications, where an AI must interpret graphic elements and then generate the code needed to reproduce or transform them.

To illustrate these capabilities, Meituan presents several demos that go beyond traditional web interfaces. LongCat-2.5 can notably generate an interactive real-time fluid simulation, build a virtual exhibition of historical objects in 3D with navigation controls, or produce miniature urban scenes. The demonstrations also cover professional uses, such as business performance analysis and the production of interactive reports based on customer feedback.

These examples show the range of intended applications, without constituting an independent validation of the model’s reliability in production environments. Marked performance gains over LongCat-2.0

The evaluations shared by Meituan show significant progress compared with the previous generation, particularly on autonomous programming tasks.

On Terminal Bench v2.1, LongCat-2.5 scores 88.4 points versus 70.8 for LongCat-2.0. The gap is even wider on Terminal Bench v4.0, where the score rises from 3.0 to 38.5. On Program Bench, which focuses on solving problems in software projects, the model improves from 2.5 to 25.5.

The reported results are also strong on more general evaluations. LongCat-2.5 scores 91.9 on BrowseComp, which measures search and navigation abilities, as well as 83.1 on MCP Atlas, which focuses on tool use. On 3DCodeBench, dedicated to 3D generation from text instructions, it reaches 55.4 points.

These figures place LongCat-2.5 close to several competing models in certain categories, without giving it overall dominance. On Terminal Bench v4.0, for example, Claude Opus 5 still maintains a significant lead with 53.9 points, while DeepSeek-V4.1-Flash scores 74.2 on DeepSWE v1.1, compared with 64.6 for LongCat-2.5.

The comparisons should also be interpreted with caution: the results presented combine measurements carried out by Meituan with scores from external publications, using evaluation setups that are not necessarily identical. Infrastructure designed for Chinese accelerators

Beyond the model’s performance, Meituan details the optimizations needed for training and running it on domestic AI accelerators.

The company says it adapted its infrastructure to the memory constraints and computing architectures available in China. Its distributed training system improves throughput per accelerator by 86%, according to its measurements, compared with a reference configuration using DDP, in supervised training tasks on long sequences.

For inference, Meituan also uses W4A4 quantization, which reduces the representation of expert weights and activations to four bits. In its tests, this solution reportedly increased initial processing throughput by 43.3% and generation throughput by 42.77%, compared with an 8-bit W8A8 configuration.

This technical dimension goes beyond cost alone. It illustrates the efforts of Chinese players to develop large models while adapting their infrastructure to a hardware environment subject to specific constraints. An API is available, but the weights are still pending

LongCat-2.5 is now accessible through the LongCat API platform. Meituan announces an allotment of 10 million tokens for new sign-ups, plus an additional 10 million for existing users. Requests benefiting from full cache reuse are also said to be free under this offer.

The documentation indicates compatibility with several agentic development environments, notably Claude Code, Codex, and Hermes, which should make it easier to integrate the model into tools already used by developers.

However, the model weights have not yet been released. The official LongCat-2.5 page mentions an upcoming availability on GitHub and Hugging Face, without specifying a timeline.

With this new generation, Meituan is seeking to bring visual understanding, software development, and autonomous execution closer together within a single model. Its internal results suggest significant progress, particularly on long tasks. It remains to be seen whether these performances will hold up when LongCat-2.5 is confronted with real-world projects that are less structured than evaluation environments.