Meituan releases LongCat-2.5 weights, with a 1.34 TB 4-bit edition

Meituan announced on November 9 that it had released the weights of LongCat-2.5, its multimodal AI model. Hugging Face lists two downloads under the MIT license. The standard repository is 3.56 TB and the quantized W4A4 repository is 1.34 TB. Hugging Face lists both as public downloads with no access approval requirement.

Meituan began offering LongCat-2.5 through its hosted API on November 2. The downloadable weights let developers run inference on compatible equipment they manage. Meituan’s reference setup for the smaller checkpoint uses 194 Huawei Atlas A2 accelerators, giving teams a concrete example of the infrastructure involved.

208 numbered shards and two additional files

The 3.56 TB and 1.34 TB figures refer to the complete repositories. The public file metadata shows 208 numbered weight shards and two separately named multimodal weight files in each. The 210 Safetensors files add up to about 3.562 TB for the standard checkpoint and 1.336 TB for W4A4. The quantized set reduces that transfer and storage volume by about 62%.

Safetensors is a format for storing the numerical arrays that make up model weights. The standard repository includes an index that maps each named tensor to its file. That mapping matters when a checkpoint is divided into hundreds of pieces. The repository also supplies configuration files and a template that defines how inputs are formatted.

Meituan describes LongCat-2.5 as a 1.6-trillion-parameter mixture-of-experts model. It says the model selects a subset of its modules for each token, activating roughly 48 billion parameters. Hugging Face’s separate tally of stored elements for the standard checkpoint is about 1.78 trillion. Meituan has provided no reconciliation of the two counts. The file list provides a direct measure of how much data a team would download.

The hardware behind the 4-bit example

The W4A4 designation refers to four-bit precision for weights and intermediate values in the model’s expert modules. The configuration assigns four bits to expert modules and eight to the linear operations selected for quantization. It exempts parts such as the input embeddings that represent tokens as vectors. The repository lists a mix of 16-bit floating-point and integer storage formats. On-disk size measures the transfer and storage burden. Inference also uses memory to process inputs and generate responses.

Meituan’s deployment guide describes loading the checkpoint on an Atlas A2 cluster. Two accelerators handle visual encoding, 64 process the initial input and 128 generate the response. That is the configuration Meituan documented for its example. The model card points to software for this NPU platform and gives no general minimum for other hardware.

Hosting the model on a team’s infrastructure gives that team control over where requests are processed and how the service is operated. In the documented setup, image processing and text generation run on separate machines. Replicating that arrangement involves connecting the visual encoder to the servers that perform generation, in addition to provisioning storage for the checkpoint.

The license and the release timeline

The repositories include MIT license files that permit copying, modification, distribution and sale, with a requirement to retain the copyright and permission notices in copies or substantial portions. The model card says the license grants no rights to Meituan’s trademarks or patents. The public files include weights, configuration, a tokenizer and code for running the model. Neither repository lists the training data.

The hosting records show repository creation on November 4 and some weight uploads on November 5. Those records establish when work appeared in the repository history. Meituan’s LongCat team publicly announced downloads at 14:05 UTC on November 9, and the model cards and documentation assets in both repositories were updated on November 10.