Close-up view of a green circuit board held by a hand, showcasing technology details
Models & capabilityBased on company claims

Meituan's LongCat-2.5 ships MIT-licensed weights and an Ascend-first deployment guide: claims checked

By Himanshu Sakre·

Photo: Arturo A / Pexels

Meituan released a 1.6 trillion parameter open-weight agent model that scores 88.4 on Terminal Bench v2.1 by its own account, with a self-hosting guide built for Huawei Ascend chips. The benchmarks are vendor numbers.

Meituan, the Chinese delivery and services giant, has released the weights of LongCat-2.5, a 1.6 trillion parameter model built for long, multi-step agent work, and published a deployment guide aimed at Huawei Ascend chips. The weights are MIT licensed. The headline benchmark score, 88.4 on Terminal Bench v2.1, comes from Meituan itself.

Two questions matter for the race: how good is it, and what does a Chinese-hardware path really prove? We checked the model card and the reporting around it.

What was released

The Hugging Face model card describes LongCat-2.5 as a native multimodal mixture-of-experts model for long-horizon agentic tasks. It has 1.6 trillion total parameters, with about 48 billion activated per token. The Decoder, reporting on November 15, 2026, adds that the API offers up to a million tokens of context and takes text, images and video, with text output.

The license is MIT for the weights. The card says it does not grant rights to Meituan trademarks or patents. That is permissive by frontier standards, and it is one reason the release matters beyond China. Our open weights versus open source explainer sets out what an MIT license on weights does and does not include.

A separate W4A4 quantized version is also listed. Meituan reports that, compared with an 8-bit baseline, it gains 43.3 percent in prefill throughput and 42.77 percent in decode throughput, with accuracy "comparable" to the unquantized model on the benchmarks it tested.

The scores

Meituan's card lists 88.4 on Terminal Bench v2.1, up from 70.8 for LongCat-2.0, 91.9 on BrowseComp, and 96.5 on MathVision with tools. The card is candid about the caveat: scores marked with an asterisk come from public sources, and unmarked scores are in-house and indicative only.

The Decoder's table puts the number in context. On Terminal Bench v2.1, it lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and DeepSeek-V4.1-Flash at 90.6. On DeepSWE, LongCat-2.5 scores 64.6 against 74.2 for DeepSeek and 72.7 for GPT-5.6 Sol. On the harder Terminal Bench v4.0 it scores 38.5, against 53.9 for Opus 5.

So the picture is mixed: near the front on the easier terminal benchmark, behind on the harder one. The Decoder also notes that newer models, Claude Opus 5.5 and GPT-6.1 Sol, are not in Meituan's table.

Close-up view of stacked rusty industrial pipes, showcasing their texture and corrosion
Racks of accelerator servers, the hardware class at the center of the LongCat-2.5 deployment guide. Photo: Engin Akyurt / Pexels

The Ascend claim

The headline in the Decoder says the model "never touched an Nvidia GPU." Be careful with that. The model card says LongCat-2.5 "can be deployed on NPU platforms" and that its W4A4 format uses native INT4 computation on domestic accelerators. The Decoder says a sample setup runs on 194 Huawei Ascend accelerators: two for images, 64 for prompt processing and 128 for generation. It also says the predecessor, LongCat-2.0, was reportedly trained on a cluster of more than 50,000 Chinese AI chips.

That is reporting, not proof of how LongCat-2.5 itself was trained. We did not find a technical report. The inference engine's public repository, SGLang-FluentLLM, states it was tested on Nvidia H800 and H20 GPUs on its main branch, and the Ascend instructions sit on a separate branch we could not open. Serving on Ascend is documented by the model card. Training exclusively on it is, for now, a claim to confirm.

Why the hardware matters

Epoch AI estimated on November 6, 2026 that Huawei's Horinger campus holds about 81,000 H100 equivalents, and that Chinese facilities get roughly 40 percent of the compute per unit of power of leading US sites. Our Horinger coverage has the detail. A model that serves well on Ascend, at 4-bit precision, is a way of making less compute go further.

Price follows. The Decoder reports promotional API pricing of about 0.27 euros per million input tokens and 1.07 euros per million output tokens, limited time. That undercuts US frontier prices, but it is a launch promotion, and we have not confirmed it against Meituan's price page.

How it was trained and how to use it

The Decoder describes the training recipe in useful detail. Reinforcement learning used real software tasks drawn from GitHub pull requests and recreated agent failures. To prevent shortcuts, the training environments had no internet access, and Git history and cached files were removed, with a language model checking solutions for tricks. Capabilities from several specialized teacher models, covering agent work, reasoning and dialog, were distilled into the final model. Images are converted at native resolution into tokens processed alongside text, rather than through a separate image encoder.

For developers, the access options are practical. The weights sit on Hugging Face under meituan-longcat, the W4A4 build is a separate listing, and a chat interface runs at longcat.ai. The Decoder says the API supports both OpenAI and Anthropic request formats and works with agent harnesses such as Claude Code and Codex, so an existing agent setup can be pointed at it with little change. The model previously ran anonymously as "Owl Alpha" on OpenRouter, according to the same report.

None of that proves quality on your tasks. The model card's own warning about in-house scores applies. A fair test is to run your real coding tasks, with your harness, against the model you use now.

Our take

This is a strong open release and a modest scoreboard story. The license, the quantization work and the Ascend serving guide are verifiable and useful. The claim of parity with US frontier models is not. We would wait for independent runs on the harder benchmark, and for a technical report that says what LongCat-2.5 was trained on, before moving it from "serious competitor" to "peer." Compare it with Kimi K3.1 Pro, the other Chinese open-weight release this month.

Frequently asked questions

What is Meituan LongCat-2.5?

A native multimodal mixture-of-experts model for long-horizon agentic tasks, with 1.6 trillion total parameters and about 48 billion active per token, per its Hugging Face model card. The weights are released under the MIT license.

Was LongCat-2.5 trained without Nvidia chips?

Not confirmed. The model card says it can be deployed on NPU platforms and uses native INT4 computation on domestic accelerators. The Decoder reports the earlier LongCat-2.0 was trained on more than 50,000 Chinese chips. No technical report states LongCat-2.5's training hardware.

How well does LongCat-2.5 score on benchmarks?

Meituan reports 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, and says unmarked scores are in-house and indicative. The Decoder lists 38.5 on the harder Terminal Bench v4.0 against 53.9 for Claude Opus 5.

Can I use LongCat-2.5 commercially?

The model card says the weights are MIT licensed, which permits commercial use. It also says the license grants no rights to Meituan trademarks or patents.

What hardware does the self-hosting guide use?

The Decoder reports a sample setup of 194 Huawei Ascend accelerators: two for images, 64 for prompt processing and 128 for response generation. The model card points to a separate NPU branch of the SGLang-FluentLLM repository.

What does the W4A4 version do?

It quantizes expert weights and activations to 4 bits. Meituan reports 43.3 percent higher prefill throughput and 42.77 percent higher decode throughput than an 8-bit baseline, with accuracy comparable on its tests.

Sources

What each one is, and whose it is.

  1. 1

    LongCat-2.5 model card, Meituan LongCat, Hugging Face (November 15, 2026)

    Model card
  2. 2

    LongCat-2.5-W4A4 model card, Meituan LongCat, Hugging Face (November 15, 2026)

    Model card
  3. Press reportIndependent of the vendor
  4. 4

    SGLang-FluentLLM, Meituan LongCat, GitHub (November 15, 2026)

    Documentation
  5. OtherIndependent of the vendor