Meituan has released LongCat-2.5, a native multimodal model built for long-horizon agentic tasks, and says it was trained entirely on Chinese-made computing hardware. The Chinese delivery and local-services group announced the model on 2 November, a day after its LongCat team published a technical blog describing a mixture-of-experts (MoE) design with 1.6 trillion total parameters and about 48 billion activated per token.
The launch matters for two reasons. It is another large Chinese model that its maker says was built without foreign accelerators, and Meituan’s own benchmark table shows it trading wins and losses with the strongest model it compares against, Claude-Opus-5. Users can try LongCat-2.5 at longcat.ai, and Meituan says developers can call it through an API that went live the same day.
In brief
- Meituan’s LongCat-2.5 is a 1.6tn-parameter multimodal MoE model, with about 48bn parameters active per token, built for agentic work such as coding.
- Meituan says it was trained entirely on domestic Chinese compute, with mid-training sequences of up to 1M tokens.
- Meituan’s own scores beat Claude-Opus-5 on BrowseComp and PaperBench but trail it on Terminal Bench v4.0 and DeepSWE. No open weights have been announced.
What Meituan says LongCat-2.5 can do
LongCat-2.5 builds on LongCat-2.0, which Meituan released in June. According to the blog, the new model adds stronger native visual understanding and better agentic coding, meaning it is designed to carry out long, multi-step software tasks rather than answer single questions.
The LongCat team says it trained the model with large-scale, long-horizon reinforcement learning on tasks such as building whole projects inside real codebases. It also says it turned real-world cases where agents failed into “reproducible, verifiable training tasks” to fill gaps in the model’s abilities.
For images, the model breaks visual input into multi-level discrete tokens, so the same backbone handles pictures and text in one token format. Meituan says this supports visual coding, creative design and 3D tasks.
The company is clear about who the model is for. Meituan’s news post quotes chief executive Wang Xing from its second-quarter earnings call on 28 August: “We will not become a token factory. Our models and AI products will be used to support our core business” and to improve the experience of users and merchants, as well as the company’s efficiency (TSN’s translation). Meituan says its LongCat models power “Miaoshou”, an AI workbench for local businesses that it says already serves chains including Guming and Yihetang.
Benchmarks: wins and losses against Claude-Opus-5
Meituan says that across several mainstream benchmarks, LongCat-2.5 scores between Claude-Opus-4.8 and Claude-Opus-5, and beats Claude-Opus-5 on some. All of these are Meituan’s own figures, not independent results. The blog table shows the mix:
| Benchmark (Meituan’s figures) | LongCat-2.5 | Claude-Opus-5 |
|---|---|---|
| BrowseComp | 91.9 | 90.8 |
| PaperBench | 82.1 | 79.1 |
| Terminal Bench v4.0 | 38.5 | 53.9 |
| DeepSWE v1.1 | 64.6 | 68.8 |
The comparison scores come from two kinds of source. Meituan says rival scores marked with an asterisk come from public technical reports, blog posts or official leaderboards, while the rest are “self-run and are indicative only”. The Claude-Opus-5 PaperBench figure is one of the self-run scores. The same table also puts other models ahead of LongCat-2.5 in places: GPT-5.6-Sol scores 83.3 on PaperBench in Meituan’s own run, and DeepSeek-V4.1-Flash, which TSN covered in October, is listed at 74.2 on DeepSWE.
Training on domestic AI accelerators
Most of the blog deals with the engineering needed to train a model of this size on Chinese hardware. Each domestic AI accelerator used has less than 80 GB of memory, the LongCat team says, so the training system was designed to save memory and keep most traffic inside high-speed “supernode” clusters. Meituan says the mid-training stage used sequences of up to 1M tokens.
The team reports that its optimised setup achieved 86% higher per-device throughput than a standard DDP (distributed data parallel) baseline in long-sequence fine-tuning tasks. For running the model, it uses W4A4 inference, which stores both expert weights and activations at 4 bits. Compared with an 8-bit W8A8 INT8 baseline, Meituan says online stress tests showed prefill throughput up 43.3% and decode throughput up 42.77%, with time to first token down 15.91% and time per output token down 19.72%. It says the 4-bit version performs comparably to the unquantised model on the tasks it evaluated.
The claim follows a similar one from iFlytek, which said in October that its Spark X2.6 model was trained only on Chinese chips. Neither Meituan’s blog nor its news post names the accelerator supplier.
What has not been announced
Meituan has not announced open weights for LongCat-2.5. As of about 08:40 GMT on 2 November, its Hugging Face organisation page listed no LongCat-2.5 model, while LongCat-2.0 and its FP8 and INT8 versions remain available there. Meituan has also not published API pricing in either announcement, or named which domestic chips it used.
Why the chip claim matters more than the scores
The benchmark table is the part of this launch most likely to be quoted, and it is also the weakest evidence. Meituan picked the tests, ran some of the rival scores itself and labels those “indicative only”. Even on its own numbers, LongCat-2.5 falls well short of Claude-Opus-5 on Terminal Bench v4.0, a test close to the long, hands-on coding work the model is pitched at. Readers should treat “between Opus-4.8 and Opus-5” as a company summary, not a ranking, much as TSN did with Moonshot’s own Kimi K3.1 Pro results.
The infrastructure detail is more telling. A 1.6tn-parameter model trained on accelerators with under 80 GB of memory each, and served at 4-bit precision on the same class of hardware, shows how much of Meituan’s effort went into software and systems design to make up for smaller, less capable chips. If its throughput figures hold up, that work matters to every Chinese lab training on the same kind of hardware.
The closed release is a change from LongCat-2.0. Keeping LongCat-2.5 to Meituan’s own site and API, for now, fits Wang Xing’s line that the models exist to serve Meituan’s own business first.
Other Chinese labs, including Z.ai with its GLM-5.3-based models, appear in Meituan’s comparison table, so independent tests on public leaderboards will be the first real check on its claims. The first to watch is Terminal Bench v4.0, where Meituan’s own figure leaves a 15.4-point gap to Claude-Opus-5.
Sources
- LongCat team, “Introducing LongCat-2.5”, 1 November 2026: https://longcat.ai/blog/longcat-2.5
- Meituan news centre, “美团发布LongCat-2.5大模型” (Meituan releases LongCat-2.5), 2 November 2026: https://www.meituan.com/news/NN261102280001343
- Meituan LongCat on Hugging Face (checked 2 November 2026): https://huggingface.co/meituan-longcat
Related stories
- iFlytek Says Its 293B Spark X2.6 Was Trained Only on Chinese Chips
- Moonshot’s Own Tests Put Kimi K3.1 Pro at 65.7% on Terminal-Bench
- DeepSeek Releases DeepSeek-V4.1-Flash-1017 Open-Weight Update
- Z.ai Launches GLM-Legal, an Agentic Legal Model Built on GLM-5.3-Flash
Image: Carl Lender / Wikimedia Commons, CC BY 2.0, modified by TSN
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
.jpg){kind=link}
