Meituan LongCat-2.5: 1.6T Open Weights Under an MIT License
LongCat-2.5 is a multimodal mixture-of-experts model from Meituan with 1.6 trillion total parameters, about 48 billion active per token, released on Hugging Face under the MIT license. It targets long agentic coding and vision tasks. It is far too large for personal hardware, and many of its benchmark comparisons use in-house runs or older rival models.
On this page
What Meituan put on Hugging Face
Meituan, the Chinese delivery and e-commerce company, released the weights for LongCat-2.5 on 9 November, a week after first introducing the model. The model card describes it as a native multimodal mixture-of-experts model for long-horizon agentic tasks, with 1.6 trillion total parameters and roughly 48 billion activated per token. It can run in a thinking mode or a non-thinking mode, and it accepts images alongside text.
The license is the headline. The card states that the weights are released under the MIT License, while adding that this does not grant rights to Meituan trademarks or patents. MIT permits commercial use, modification, and redistribution with only a notice requirement, which makes it simpler than open-weight licenses that carry revenue thresholds or usage policies.
One detail is worth flagging. The card's text says 1.6 trillion parameters, but the Hugging Face metadata for the repository lists 1.8 trillion. VentureBeat noticed the same gap and advises sizing any deployment from the actual files.
Open does not mean local at this size
The 48 billion active figure can make LongCat-2.5 sound manageable. It is not, at least for personal machines. A mixture-of-experts model only computes with a fraction of its weights for each token, but all of the weights still have to be stored and reachable. VentureBeat's illustrative arithmetic puts 1.6 trillion parameters at about 3.2TB in 16-bit precision, while noting that this is not the real checkpoint footprint.
Meituan's own deployment notes point at datacenter hardware. The card says the model can be deployed on NPU platforms through a Meituan fork of SGLang, and that it applies 4-bit quantization to expert weights and activations so that it can compute natively in INT4 on domestic AI accelerators. A separate 4-bit repository sits alongside the main one on Meituan's Hugging Face page.
The card we read gives no memory or GPU requirement, and it does not mention GGUF, llama.cpp, or other tools that hobbyists rely on. Hugging Face also shows no inference provider hosting the model yet. For most people, the practical route today is Meituan's own API, which VentureBeat reports at a limited-time $0.30 per million input tokens and $1.20 per million output tokens.
How to read the benchmark table
The improvement over the previous version is large. Meituan reports 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, and 64.6 on DeepSWE v1.1, up from 9.7. On BrowseComp, its 91.9 is the highest score in the card's table.
Against rivals, the picture is mixed. On Terminal Bench v2.1, DeepSeek-V4.1-Flash (90.6) and Claude Opus 5 (89.1) both sit above it. On the harder Terminal Bench v4.0, LongCat-2.5 scores 38.5, behind GLM-5.3 at 41.8 and Claude Opus 5 at 53.9. On DeepSWE v1.1, DeepSeek-V4.1-Flash leads with 74.2.
Two caveats apply. First, the card says scores without an asterisk come from Meituan's in-house runs and are indicative only, while the others are cited from public sources. Second, the comparison set skips the newest closed models. VentureBeat points to leaderboard scores of 58.2 for Claude Opus 5.5 and 64.8 for GPT-6.1 Sol on Terminal Bench v4.0, well above LongCat's 38.5, and calls Meituan's framing a sub-frontier case.
Why it still matters for private AI
Nobody will run LongCat-2.5 on a laptop, but permissively licensed weights of this size change what organisations can do. A hospital, bank, or government lab with serious hardware can host the model on its own infrastructure, keep prompts and documents inside its own network, and modify the model without asking permission. Cloud providers can also offer it competitively, which tends to push prices down.
There are limits. VentureBeat notes that self-hosting does not remove supply-chain or procurement questions around Chinese-developed models, although it found no US federal restriction naming Meituan or LongCat-2.5 as of 9 November. Meituan also says the model has not been evaluated for every downstream use and may perform unevenly across languages.
The thing to watch is what the community does next. MIT terms allow anyone to distill, prune, or fine-tune these weights into smaller models, and that is usually how releases of this scale eventually reach hardware a person can own.