Meituan released LongCat-2.5 model weights under an MIT license

Meituan has released the weights for LongCat-2.5 under an MIT license, VentureBeat reported at 3:04 p.m. PT on November 9, 2026. VentureBeat said the Chinese food-delivery and e-commerce company published the checkpoint on Hugging Face, one week after introducing the model. A Meituan technical-team post dated November 10, 2026 also says the company has formally open-sourced LongCat-2.5 under the MIT license, allowing commercial inference, deployment and further training without additional authorization. This article summarizes those company and press accounts. It is not a laboratory evaluation of the model.

Real, reported: based on the public sources cited below. Nothing added or dramatized. Sources checked November 11, 2026, at 7:15 AM ET (EST). The release, license and architecture claims are company statements. Benchmark scores discussed below are company-reported, or are described by VentureBeat as coming from Meituan's materials or from public leaderboards under mixed test setups. This is a company claim, not a peer-reviewed paper, and neither source names a scientific journal.

What Meituan says it released

Meituan's November 10 post describes LongCat-2.5 as a native multimodal model aimed at long-running agent tasks, including complex codebase work, full project construction, visual coding and professional office work. It says the model has 1.6 trillion total parameters, activates about 48 billion parameters per token, and supports a one-million-token context. VentureBeat similarly reported about 48 billion parameters active per token out of 1.6 trillion total, citing Meituan's model card, and noted that a Hugging Face snapshot displayed 1.8T in its metadata. VentureBeat advised readers to size any deployment from the actual files and serving format rather than from that arithmetic alone.

Meituan says the model represents visual signals as multi-level discrete tokens that the main model processes together with text, rather than as a bolted-on vision module. It also describes a W4A4 inference approach that quantizes expert weights and activations to 4 bits, and it says evaluated downstream-task performance remained comparable to the version before quantization. Those efficiency claims are Meituan's. We have not independently measured speed, memory use or quality after quantization. The same Meituan post also says the company trained and fine-tuned the model in a long-sequence setting on domestic accelerators with less than 80 GB of memory per card, and it reports that one FSDP configuration delivered 86% higher per-card throughput than a DDP baseline on a long-sequence supervised fine-tuning task. That 86% figure is a company-reported engineering result on a stated baseline, not an independent hardware benchmark.

The company says it chose the MIT license so the model can be freely used, modified and redistributed, subject to the license terms. VentureBeat likewise reported that the MIT license on the weights allows commercial use, modification and redistribution subject to notice requirements, and that the hosted API is a separate purchase decision.

How the reported scores should be read

VentureBeat summarized comparisons in Meituan's evaluation, including scores against the model's predecessor: 88.4 versus 70.8 on Terminal Bench v2.1, 64.6 versus 9.7 on DeepSWE v1.1, and 58.3 versus 33.1 on a SWE Atlas codebase question-answering test. Those figures are company-reported comparisons with LongCat 2.0, not results we reproduced.

VentureBeat also cautioned that Meituan's rival comparisons are more mixed, and that some proprietary reference points were older models whose successors were already available when LongCat-2.5 was introduced. It wrote: "That is a reading of the evidence Meituan chose to publish. The choice alone cannot establish how LongCat ranks against today's best models." VentureBeat further noted that some leaderboard figures for newer models use different tools and test setups, and that a dash in its table means that comparison had no verified result to show. We repeat that caution rather than treating any row as a definitive ranking.

Meituan's own post emphasizes training changes meant to support long tasks, including more long-running traces in the training data and constraints intended to reduce shortcut-taking behavior. It does not, in the text cited here, present those design choices as a peer-reviewed finding.

Limits and practical caveats in the coverage

VentureBeat's account reports a company's weight release and adds caveats for buyers. It is not an independent audit. It notes that operating the full model on a private network is a different calculation from trying a hosted API, and that teams should test whether a serving runtime can load the checkpoint and preserve the model's visual and tool-use behavior. It also reported that it found no publicly disclosed U.S. federal procurement or use restriction that specifically names Meituan or LongCat-2.5 as of November 9, while adding that rules can change quickly. That is VentureBeat's legal-policy observation, not a clearance or a ban, and it is not legal advice.

Our advice (general guidance)

Treat LongCat-2.5 as a newly published company model, not as a certified or independently scored product. If you try the weights or a hosted interface, use public or synthetic examples first. Do not upload customer data, private source code, or internal documents until you have read the license, the host's retention terms and any rules that apply to your work. Company-reported benchmark tables are a starting point for a trial, not a substitute for testing the tasks you actually need, with the same tools and time limits for every model you compare.

Be cautious of third-party sites or messages that claim to offer "official LongCat weights" or a free API key. Use the project pages Meituan itself publishes, and be wary of anyone asking you to disable security tools or to pay for a download of an open-source checkpoint.

These are editorial precautions, not individualized engineering, procurement or legal advice.

Sources

Published November 11, 2026 by FraudMeNot Team. Label: Real, reported.

Disclaimer. General education only, not legal, financial or security advice. No guarantee. Consult a qualified professional. News summaries are based on the sources listed and may change as more is reported.

General education only, not legal, financial or security advice. Not affiliated with the organizations discussed. No sponsorship or affiliate links. Information may change as further evidence becomes available.