LongCat-2.5

Meituan's 1.6T multimodal mixture of experts with 48B active, a 1M context and the plain MIT license. Documented only for NPU servers.

In short. LongCat-2.5 is a huge AI model from the Chinese delivery platform Meituan that anyone may download and use under the MIT license. It reads text, images and video, and it is built for long coding and office tasks that take many steps. Running it takes a cluster of data-center accelerators, and Meituan documents deployment only for NPU servers.

Summary

Meituan introduced LongCat-2.5 on 2 November 2026 and released the weights on 9 November. It is a native multimodal mixture-of-experts model with 1.6 trillion parameters, about 48 billion of them active per token, built on LongCat-2.0 from July. Instead of adding a separate vision module, it turns image features into multi-level discrete tokens that the main network handles like text, and it also takes video. Meituan says it supports a one-million-token context, and it is a hybrid model whose thinking mode can be switched on or off per request.

The model card describes reinforcement learning on real engineering tasks and reproducible agent failures, with several specialist teacher models guiding the final model. It also applies W4A4 quantization to expert weights and activations, designed for native INT4 computation on domestic Chinese AI accelerators. The BF16 files total about 3.56 TB and the W4A4 build about 1.34 TB. In Meituan's own table the model scores 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, against 88.3 and 91.2 for Kimi K3 and 90.6 and 91.5 for DeepSeek-V4.1-Flash. Rival scores marked with an asterisk are copied from public sources, and the others, including Meituan's own, are in-house runs that Meituan calls indicative.

The weights are under the MIT license, which Meituan says allows commercial inference and further training without extra authorization. The license grants no rights to Meituan trademarks or patents. The model card documents deployment only on NPU platforms, through Meituan's SGLang-FluentLLM fork, and gives no GPU instructions. If you do not want to host it, Meituan offers the model through its API platform and the LongCat chat website.

Good for

  • Labs and companies with a large accelerator cluster that want a frontier-scale model under the plain MIT license
  • Long-horizon coding agents, terminal work and browsing research
  • Visual coding, such as creative design and 3D tasks, and reading screenshots and video alongside tool use
  • Studying discrete visual tokens and W4A4 serving of a trillion-parameter mixture of experts

Watch out for

  • The BF16 files total about 3.56 TB and even the W4A4 build is about 1.34 TB, so it is out of reach for personal hardware
  • The model card documents NPU deployment only, through a Meituan fork of SGLang, with no GPU or GGUF instructions
  • Meituan announces a 1M-token context, but config.json lists 524,288 positions, so test long inputs before relying on them
  • The card says 1.6T parameters while the Hugging Face metadata shows about 1.8T, so size a deployment from the actual files
  • Benchmark numbers come from Meituan, and unmarked scores are in-house runs that Meituan calls indicative

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the LongCat-2.5 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
LongCat-2.5 BF16 1.6T 48B active 3.68 TB 1.84 TB 1.1 TB GPU server
LongCat-2.5-W4A4 INT4 1.6T 48B active 3.68 TB 1.84 TB 1.47 TB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

See also

Frequently asked questions

What is LongCat-2.5?
LongCat-2.5 is a huge AI model from the Chinese delivery platform Meituan that anyone may download and use under the MIT license. It reads text, images and video, and it is built for long coding and office tasks that take many steps. Running it takes a cluster of data-center accelerators, and Meituan documents deployment only for NPU servers.
Can I use LongCat-2.5 commercially?
Yes. LongCat-2.5 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is LongCat-2.5 open source?
LongCat-2.5 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run LongCat-2.5 on my own computer?
Not on a desktop. The smallest variant, LongCat-2.5, needs about 1.1 TB at 4-bit precision, which means a GPU server with several data-center cards.
When was LongCat-2.5 released?
LongCat-2.5 was released by Meituan on November 9, 2026.
How long a context window does LongCat-2.5 have?
LongCat-2.5 supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Kimi K3.1 Pro Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
DeepSeek V4.1 Flash DeepSeek 552B 16B active 1M MIT (Permissive) 561 GB GPU server
Hy4 preview Tencent 770B 49B active 1M Apache 2.0 (Permissive) 531 GB GPU server