Meituan releases LongCat-2.5 weights under MIT license

Meituan has released LongCat-2.5 model weights under the MIT license, putting the newer multimodal model’s checkpoint and configuration files directly in developers’ hands.

The repository file tree lists a 3.56 TB checkpoint, an index and 208 numbered Safetensors shards. Its downloadable configuration enables visual, visual-quantizer and video components and includes identifiers for image and video tokens.

Meituan is using the standard MIT license. That differs from another recent Hugging Face open-weight release, whose terms limited commercial model-as-a-service use.

Meituan’s model card describes LongCat-2.5 as a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion parameters activated for each token. A mixture-of-experts model routes each token through only part of the network instead of using every parameter for every request.

The company says LongCat-2.5 combines visual understanding with agentic coding for complex codebases, full-project builds, creative design and 3D tasks. The available evidence does not independently establish those capabilities.

The model card also reports scores of 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1 and 78.6 on Vision2Web. It says unstarred scores were measured in-house and are indicative only.

The repository history records weight-shard upload commits on November 4 and 5, followed by a model-card update on November 9. The available pages do not establish exactly when the repository or weights first became publicly visible.

The repository’s license file contains the standard MIT license with a 2026 Meituan copyright notice. The model card says the license does not grant rights to Meituan trademarks or patents.