Meituan releases LongCat-2.5 weights: 1.6T parameters, MIT license and multimodality

China's Meituan, best known for food delivery, continues to make a serious push into open-source AI. The LongCat team has released LongCat-2.5's weights on Hugging Face under the MIT license. It is one of the most permissive licenses: the model can be fine-tuned, integrated into products and sold without separate agreements.

What's inside

LongCat-2.5 uses a Mixture-of-Experts architecture. It has 1.6 trillion parameters in total, but only around 48 billion activate for each token. The idea is simple: think of the model as a large office with hundreds of specialists, where only the few needed for a particular question respond. It delivers quality close to that of a huge model while requiring roughly the per-token computation of a midsize model.

The second defining feature is native multimodality. Vision is not bolted on through a separate encoder; it was trained jointly with the rest of the model. Its developers claim strong visual understanding and agentic coding in a single model. Examples include complex coding projects, creative design and 3D tasks.

Why this matters in practice

Consider a typical task: a designer sends a screenshot of a mockup, and an agent needs to turn it into a working interface, run it, inspect the result and fix the layout. This usually involves connecting a vision model to a coding model, with context lost in the handoff. If one model can see the screen and write code, the "inspect, fix, verify" loop becomes shorter and more reliable.

For companies that cannot send code and mockups to external APIs, this also offers a chance to run a capable agent on their own infrastructure. The MIT license removes the legal questions that often get in the way when models come with custom terms.

Limitations

First and most importantly, the announcement contains no benchmarks at all. Claims of "stronger" vision and coding remain unsupported, and the community still needs to compare the model with other open models. Until independent measurements are available, these claims should be treated as marketing.

Second, those 1.6 trillion parameters need to be stored somewhere. Even with only 48B active, the entire model must reside in memory, requiring around 1.6 TB in 8-bit format. That means a cluster of several GPU servers, rather than a workstation. In practice, large teams or cloud providers will be able to run it on their own hardware; everyone else will be waiting for quantized versions and hosting through third-party APIs.

Third, tool support, context length and reliability across long agent workflows remain unclear. These are precisely the areas where most models that shine in tests break down in real projects.

What to do

If you are building coding agents or working on image-to-code tasks, add LongCat-2.5 to your shortlist and test it on your own workloads, rather than relying on other people's benchmarks. If independence from closed APIs matters to you, watch for quantized versions and support in vLLM and SGLang. And if you are a product manager, factor in the broader trend: open models from China with MIT licenses and trillion-parameter scale have become commonplace, putting pressure on closed providers' prices.