Meituan announced LongCat-2.5 on November 2, 2026 and open-sourced its weights on November 10 under the MIT licence. It follows LongCat-2.0, which Meituan says it trained and served on a cluster of Chinese-made accelerators. This article covers what is new in 2.5, what carries over from 2.0, and what Meituan’s own posts and repositories do and do not show.
Meituan opens LongCat-2.5
Meituan announced LongCat-2.5 on November 2, 2026, describing it as a native multimodal model for long-horizon agentic tasks (translated from Chinese by TechiesJournal). Eight days later, Meituan’s November 10 post said the model was open-sourced under the MIT licence, with no extra authorisation needed for commercial inference deployment or further training (translated). The Hugging Face repositories for LongCat-2.5 and its W4A4 quantised build were created on November 4 and last modified on November 10, which matches the announce-then-open pattern Meituan describes.
LongCat-2.5 keeps the headline size of its predecessor: Meituan’s model card and both November posts state 1.6 trillion total parameters with about 48 billion activated per token. The Hugging Face API separately counts about 1.78 trillion stored tensor parameters across the repository’s BF16 and F32 weight files, a figure Meituan does not reconcile with its own 1.6 trillion figure, the same gap found for LongCat-2.0.
What changed from LongCat-2.0
LongCat-2.0 launched first. Meituan announced it on June 30, 2026, published its weights on Hugging Face in the first days of July, and described the open-source release and its serving work in a July 12 technical post. That model is a text-only mixture-of-experts system with the same 1.6 trillion total and roughly 48 billion activated parameters, a dynamic activation range of 33 to 56 billion through what Meituan calls zero-computation experts, and advertised support for context around one million tokens (LongCat-2.0 model card). For readers comparing LongCat with the wider move toward domestically built AI infrastructure, TechiesJournal’s article on why China is building a full AI stack covers the broader pattern LongCat-2.0 and LongCat-2.5 both sit inside.
LongCat-2.5 adds three things LongCat-2.0 does not have. First, native multimodality: the model card describes visual signals converted into discrete tokens and processed by the same backbone as text, built on Meituan’s earlier LongCat-Next work, rather than a separate vision module bolted onto a text model. Second, W4A4 inference, which quantises expert weights and activations to 4 bits and runs them on native INT4 matrix units, a format LongCat-2.0 does not use. Third, a narrower public deployment story: the LongCat-2.5 model card states only that it “can be deployed on NPU platforms,” while LongCat-2.0’s card also documents a GPU route through an SGLang cookbook. Meituan’s November 2 post (translated) also says the long-horizon agentic coding gains in 2.5 come from a wider reinforcement-learning pipeline and more real-world failure trajectories converted into reproducible training tasks, built on top of LongCat-2.0’s existing expert-teacher training setup rather than replacing it.
| Attribute | LongCat-2.0 | LongCat-2.5 |
|---|---|---|
| Total parameters (Meituan) | 1.6 trillion | 1.6 trillion |
| Average activated parameters (Meituan) | About 48 billion, range 33 to 56 billion | About 48 billion |
| Modality | Text only | Native multimodal (image and text) |
| Advertised context | Around 1 million tokens | Around 1 million tokens |
| Quantised inference format | BF16, FP8, INT8 | BF16, plus a separate W4A4 (4-bit) repository |
| Documented deployment targets | NPU (Ascend Atlas A2 guide) and GPU (SGLang cookbook) | NPU only, per the model card |
| Licence | MIT | MIT |
| Announced | June 30, 2026 | November 2, 2026 |
| Weights published | First days of July 2026 | November 10, 2026 |
Source: LongCat-2.0 model card and LongCat-2.5 model card, read November 12, 2026.
What Meituan says about training and inference on Chinese accelerators
Meituan’s claim about LongCat-2.0 is specific and, on the public record, consistent across its own channels. Its June 30 post (translated) calls LongCat-2.0 “the industry’s first trillion-parameter model to complete full-pipeline training and inference on a 50,000-card domestic compute cluster,” trained from scratch. Its July 12 post makes the inference half of that claim again on its own: “the industry’s first trillion-parameter model to complete inference on a 50,000-card domestic compute cluster” (translated). These are Meituan’s own statements, repeated by the company itself, not an outside observer’s conversion of an inference figure into a training figure.
For LongCat-2.5, Meituan’s November 10 post (translated) says both training and supervised fine-tuning used long-sequence training on Chinese accelerators, describing per-card memory under 80 GB as the constraint its parallelism and memory-sharding work had to solve, and reporting an 86 percent higher per-device throughput than a standard data-parallel baseline on its long-sequence fine-tuning tasks. The LongCat-2.5 model card repeats the 86 percent figure in English. Neither post gives a card count for the LongCat-2.5 training run comparable to the 50,000-card figure Meituan gives for LongCat-2.0.
The public material does not independently establish cluster throughput, reliability, cost, or how much of this work could run on other hardware. No training code for either model was found in Meituan’s repositories (see the table below), so none of this can be reproduced outside Meituan.
| What Meituan claims | What is public |
|---|---|
| LongCat-2.0 trained and served on a 50,000-card domestic cluster (June 30 and July 12 posts) | Not independently verified. No training code published |
| LongCat-2.5 trained and fine-tuned on Chinese accelerators with per-card memory under 80 GB (November 10 post) | Not independently verified. No training code published, and no card count given for 2.5 |
| W4A4 inference raises prefill throughput 43.3% and decode throughput 42.77% over a W8A8 baseline (LongCat-2.5 model card and November 10 post) | Meituan’s own benchmark, not independently reproduced |
| Both models run on Chinese-made accelerators | NPU deployment guide names Huawei Ascend Atlas A2 systems. Meituan’s posts say only “domestic” or “Chinese-made” chips and do not name a vendor themselves |
Source: Meituan’s June 30, July 12 and November 10 posts, and the LongCat-2.5 model card, read November 12, 2026.
Licence and how to get the weights
LongCat-2.5’s weights are released under the MIT licence. The LICENSE file is the standard MIT text, and the model card adds that the licence “does not grant any rights to use Meituan trademarks or patents.” LongCat-2.0 carries the identical MIT text. MIT permits commercial use, modification and redistribution.
The weights are split across two Hugging Face repositories: LongCat-2.5, stored in BF16 across 194 safetensors shards, and LongCat-2.5-W4A4, the 4-bit quantised build. Neither repository is gated. As of November 12, 2026, no meituan-longcat/LongCat-2.5 repository existed on GitHub, and Meituan’s English technical blog for 2.5 still showed its GitHub link as not yet published. This matches LongCat-2.0’s own GitHub repository, which holds only a README, a licence file and figures, no training code.
For teams weighing whether an open-weight Chinese model fits their stack, TechiesJournal’s earlier look at DeepSeek, GLM and Kimi open-weight models covers the same licence-and-usability questions for three other model families.
Which chips, and can it run on GPUs
Neither the November 2 nor the November 10 post names a specific accelerator for LongCat-2.5. Both describe only “domestic” or Chinese-made AI accelerators and chips (translated). The one place a vendor is named is a deployment guide inside Meituan’s SGLang-FluentLLM repository, on its longcat-2.5-npu branch, which targets Huawei Ascend Atlas A2 Pod systems and describes a 194-card example cluster for the W4A4 build, split as 2 cards for a visual encoder, 64 for prefill and 128 for decode. Meituan does not say whether the chips in its training cluster are the same Ascend hardware named in this inference guide.
The LongCat-2.5 model card states only that the model “can be deployed on NPU platforms.” Unlike LongCat-2.0, whose model card also documents a GPU deployment route through an SGLang cookbook, LongCat-2.5’s card gives no GPU deployment instructions. A team planning to serve LongCat-2.5 on GPUs has no documented path from Meituan today. Readers following the broader buildout of China’s AI accelerator and data centre industry can see more of that context in TechiesJournal’s survey of China’s AI infrastructure, chips, networks and cloud.
| Item | Status on November 12, 2026 |
|---|---|
meituan-longcat/LongCat-2.5 on Hugging Face | Public, MIT licence, BF16, created November 4, last modified November 10 |
meituan-longcat/LongCat-2.5-W4A4 on Hugging Face | Public, MIT licence, 4-bit quantised, created November 4, last modified November 10 |
meituan-longcat/LongCat-2.5 on GitHub | Does not exist. Meituan’s English blog still showed the GitHub link as not yet published |
| NPU deployment guide | Published, in the SGLang-FluentLLM repository’s longcat-2.5-npu branch, names Huawei Ascend Atlas A2 |
| GPU deployment guide | None found for LongCat-2.5 |
| Hosted API listing (OpenRouter) | Lists meituan/longcat-2.0 only. No LongCat-2.5 listing found as of this read |
Source: Hugging Face and GitHub, read November 12, 2026.
What a reader can do today
A team evaluating LongCat-2.5 can download the weights now, under an MIT licence that permits commercial use and further training without asking Meituan first. A team planning to self-host it should plan for NPU hardware, specifically the Huawei Ascend Atlas A2 platform documented in Meituan’s own deployment guide, since no GPU deployment instructions for LongCat-2.5 have been published. A team citing Meituan’s training-and-inference claims, on either model, should attribute them to Meituan by name and post date, because no training code or independent reproduction of the 50,000-card LongCat-2.0 claim, or the per-card memory figures Meituan gives for LongCat-2.5, was found. A team reading any benchmark number from either model card should treat it as Meituan’s own figure, since the cards themselves label unmarked scores as in-house and only starred scores as cited from other publishers’ reports.
References
- Meituan: LongCat-2.5 announcement, November 2, 2026 (Chinese)
- Meituan: LongCat-2.5 open-source announcement, November 10, 2026 (Chinese)
- LongCat-2.5 model card on Hugging Face
- LongCat-2.5-W4A4 model card on Hugging Face
- LongCat-2.5 NPU deployment guide, SGLang-FluentLLM repository
- Meituan: LongCat-2.0 announcement, June 30, 2026 (Chinese)
- Meituan: LongCat-2.0 technical post, July 12, 2026 (Chinese)
- LongCat-2.0 model card on Hugging Face
- LongCat model organisation on GitHub
- OpenRouter: Meituan LongCat-2.0 listing
Sources were read on November 12, 2026. The Chinese-language Meituan posts were translated by TechiesJournal, faithfully but not by a certified translator. Meituan’s figures and claims are its own and were not independently tested.