Meituan released the weights of LongCat-2.5 on 9 November, publishing the model on Hugging Face under the MIT licence a week after introducing it, VentureBeat reported. The release gives developers a way to run the Chinese food delivery and e-commerce company’s vision-capable coding model themselves, rather than only through Meituan’s hosted service.
Key points
- Meituan has published LongCat-2.5’s weights on Hugging Face under the MIT licence.
- The model has 1.6 trillion parameters, with about 48 billion active for each token it processes.
- Meituan’s Terminal Bench v2.1 score rose from 70.8 for LongCat-2.0 to 88.4 for LongCat-2.5.
Hugging Face release adds local deployment
The weights are the numerical settings a model learns during training. Publishing them lets others load and run the model on their own infrastructure, rather than sending every request to Meituan’s API. VentureBeat reported that the weights could support tasks combining screenshots or video with code, while running the entire model within an organisation’s restricted systems would require an assessment unlike use of Meituan’s hosted service.
The 9 November publication followed an earlier period of hosted access. A LongCat-2.5 preview had been available on Meituan’s API platform since 25 September, Pandaily reported in its account of the model’s introduction. An earlier AI Affairs report on LongCat-2.5 covered its release for long-running coding tasks. The weights publication changes how developers can obtain and deploy it, rather than introducing a different model.
The MIT licence gives developers a permissive basis for building on the published weights. That matters for work which needs a model to inspect an image, navigate a codebase and revise software over several steps: access to the weights allows the model to be run outside Meituan’s hosted API. It does not make the full model small. Its design reduces the amount of computation used for each piece of text, but its complete set of weights still has to be accommodated for private deployment.
LongCat-2.5 uses 48 billion parameters per token
LongCat-2.5 has 1.6 trillion parameters in a mixture-of-experts architecture, of which about 48 billion are active for each token. Think of a large reference library: every volume belongs to the collection, but a particular question calls for only a selection of them. The active figure describes the part used to process a token; the larger figure describes the model whose weights have now been released. The distinction is important when considering both the work done during a request and the size of the model being deployed.
Meituan represents visual information as discrete tokens at multiple levels, allowing the model’s main processing system to handle it alongside text. The company has paired that approach with training aimed at long sequences of actions in coding tasks. In the training account reported by Pandaily, Meituan says it expanded the data covering extended task trajectories and used reinforcement learning for engineering work across complex codebases. It also turns failed agent attempts into tasks that can be reproduced and checked during further training.
Fixing a broken checkout page could require comparing a screenshot with the code behind it, changing the relevant part and checking the rendered result. A model able to process the image and work through the code could attempt those steps together. The release would allow that attempt to be run on infrastructure other than Meituan’s API, while the published weights and the training approach alone would not make a proposed fix reliable.
Meituan also says it optimised W4A4 inference on domestic chips, using four-bit values for both weights and activations, the intermediate numbers produced as the model runs. Lower-precision numbers require less space to represent than higher-precision ones, an engineering consideration for a model with so many total parameters. Meituan says it worked on keeping training and inference numerically consistent, as well as on the long-context processing used during reinforcement learning. The account describes an optimisation effort, not a measurement of how the released weights perform under production load.
Meituan reports 88.4 on Terminal Bench
On Meituan’s own Terminal Bench v2.1 evaluation, LongCat-2.5 scored 88.4 against 70.8 for LongCat-2.0. The company’s table also gives the newer model 51.8 on AutomationBench against 15.8 for its predecessor. Those are results from the company’s benchmark table for named tests, rather than measurements of how often the model completes a developer’s work successfully. Pandaily reported that the figures had not been independently verified.
The results vary by task. Meituan’s table lists 55.4 on 3DCodeBench and 78.6 on Vision2Web for LongCat-2.5, while its Program Bench score is 25.5 against 37 for the top entry in that comparison. VentureBeat reported that terminal operations, repository fixes and turning an image into a web interface are distinct jobs. For a developer choosing what to test, the model’s improvement on one benchmark cannot stand in for a result on another.
The model’s compatibility with existing coding environments provides another route for trying those tasks. Meituan’s change log lists Claude Code, Codex and Hermes among the agent harnesses LongCat-2.5 works with, while its platform documentation includes setup guides for OpenCode, Cline, Kilo Code, CodeBuddy and WorkBuddy.
