Meituan released LongCat-2.5’s weights on November 9, 2026, a week after its introduction, VentureBeat reported. The checkpoint is available on Hugging Face under MIT terms, giving researchers a downloadable model to examine alongside its configuration and deployment software.
What Meituan released
Meituan’s model card describes LongCat-2.5 as a native multimodal mixture-of-experts model combining visual understanding with agentic coding. It lists 1.6 trillion total parameters, with approximately 48 billion activated per token. These are the developer’s specifications.
The release makes the checkpoint an open weights model. Its MIT license permits use, modification and redistribution, subject to retaining the copyright and permission notices in copies or substantial portions. The text also disclaims warranties. This gives university teams a concrete model weight licence to assess when planning reuse.
Access rights and deployment capacity are separate
The primary release materials document different parts of the research infrastructure. Reading them together distinguishes permission to reuse a checkpoint from the computing environment described for running it.
| Release material | What it establishes | Research implication |
|---|---|---|
| Model repository and MIT license | The card explicitly applies MIT terms to the model weights. | Reuse has an identified license and notice condition. |
| Published configuration | Visual and video configuration sections are enabled. | The release includes configuration evidence for its multimodal design. |
| SGLang-FluentLLM NPU example | The W4A4 example uses 194 Atlas A2 accelerators: two for encoding, 64 for prefill and 128 for decoding. | The documented example is an accelerator-cluster deployment, relevant to institutional capacity planning. |
That hardware count describes the published example, rather than a universal minimum for every implementation. For university computing services, this distinction matters: downloadable weights establish access, while the runtime and infrastructure determine whether a particular deployment is feasible.
What the benchmark claims mean for researchers
Meituan reports coding, general-agent and visual evaluations. Its card identifies unmarked scores as in-house measurements that are indicative only; starred scores come from external reports or leaderboards. Those results describe the developer’s evaluation evidence, rather than independently established performance in a university workflow.
The card also says the model has not been comprehensively evaluated for every downstream application. The research opportunity is therefore a checkpoint that can be studied and preserved with its configuration. A reproducible AI experiment would still depend on documenting the runtime and evaluation conditions. For research offices, the release separates three questions: the permission attached to the weights, the resources required for an implementation, and the evidence supporting its intended use.
Sources
- Meituan: LongCat-2.5 model card — checked 2026-11-14.
- Meituan: LongCat-2.5 MIT license — checked 2026-11-14.
- Meituan: LongCat-2.5 configuration — checked 2026-11-14.
- Meituan: SGLang-FluentLLM NPU deployment example — checked 2026-11-14.
- VentureBeat: release date report — checked 2026-11-14.
