Meituan's LongCat-2.5: 1.6T parameters, vision and the plain MIT license
Meituan released LongCat-2.5 as open weights on 9 November: a 1.6T mixture of experts with 48B active, image and video input, a 1M-token context claim and the MIT license.
Meituan released the weights of LongCat-2.5 on 9 November 2026, a week after it announced the model. It is a 1.6-trillion-parameter mixture-of-experts model with about 48 billion parameters active per token. It reads images and video as well as text, and it ships under the MIT license.
What is new
LongCat-2.5 builds on LongCat-2.0 from July. Rather than bolt on a separate vision module, it turns image features into multi-level discrete tokens that the main network handles like text. One checkpoint has a thinking mode and a non-thinking mode, chosen per request. Meituan announces a one-million-token context window, but config.json in the repository lists 524,288 positions, so test long inputs before you rely on them.
The jump over the previous version is large in Meituan’s own tests. VentureBeat, summarizing Meituan’s evaluation, reports 88.4 on Terminal Bench v2.1 against 70.8 for LongCat-2.0, and 64.6 on DeepSWE v1.1 against 9.7. Against current rivals, Meituan’s table shows 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, next to 88.3 and 91.2 for Kimi K3 and 90.6 and 91.5 for DeepSeek-V4.1-Flash. Rival scores marked with an asterisk are copied from public sources, and the rest are in-house runs that Meituan calls indicative.
Running it
The BF16 files total about 3.56 TB. The W4A4 build keeps expert weights and activations in 4 bits for native INT4 computation on Chinese AI accelerators, and it is still about 1.34 TB, more than twice the 640 GB of a server with eight 80 GB GPUs. The model card documents deployment only on NPU platforms, through Meituan’s SGLang-FluentLLM fork, and gives no GPU or GGUF instructions. The Hugging Face metadata shows about 1.8T parameters while the card says 1.6T, so size a deployment from the actual files. Meituan also sells access through its API platform, and VentureBeat reports a limited-time price of US$0.30 per million uncached input tokens and US$1.20 per million output tokens.
License
The weights are under the MIT license, which Meituan says allows commercial inference and further training without extra authorization. It grants no rights to Meituan’s trademarks or patents. There is no territory limit, revenue cap or hosting clause. That sets it apart from Kimi K3.1 Pro, GLM-5.3 and Mistral Large 4, whose licenses add conditions for large hosting businesses. Hosting clauses in open licenses compares what each one asks.
Why it matters
LongCat-2.5 is a frontier-size model that a company may modify and host without asking anyone, and it handles screenshots and video in the same agent loop as code. Very few teams can run it, because the only documented path is a Chinese NPU cluster, so the practical value for most readers is as a hosted model and as a sign of where open weights are heading. All of the scores come from Meituan, and VentureBeat points out that newer proprietary models than the ones in Meituan’s comparison were already on sale at launch.
If you have one server rather than a cluster, the best open-source LLMs and coding and agents lists show models that fit, and the multimodal models list covers the others that read images.
Sources
- LongCat-2.5 technical blog (Meituan LongCat) longcat.ai
- LongCat-2.5 model card (Hugging Face) huggingface.co
- LongCat-2.5 LICENSE (MIT) huggingface.co
- LongCat-2.5 is open source (Meituan Tech Blog, 10 Nov 2026, in Chinese) tech.meituan.com
- Meituan releases the open weights of LongCat-2.5 (VentureBeat, 9 Nov 2026) venturebeat.com
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| LongCat-2.5 Meituan | 1.6T 48B active | 1M | MIT (Permissive) | 1.1 TB GPU server | |
| Kimi K3 Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| DeepSeek V4.1 Flash DeepSeek | 552B 16B active | 1M | MIT (Permissive) | 561 GB GPU server | |
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster |