- Meituan released the weights of LongCat-2.5, a 1.6-trillion-parameter agent model, under the MIT license. Its self-hosting guide covers Huawei Ascend hardware only, according to The Decoder.
- Meituan's own benchmark scores put the model near leading Western models on some coding tests and well behind on others.
- Leaders should treat the hardware path, licence terms and independent test results as separate questions before choosing an agent model.
AI purchases often focus on the model and leave the chips underneath as an afterthought. A release from Meituan, the Chinese food-delivery company, puts that habit in question. The model deserves attention on its own terms. The more interesting part is the stack beneath it.
What Meituan released
Meituan's LongCat team published the weights of LongCat-2.5 on Hugging Face. It describes the model as built for long agentic tasks, such as complex codebase development and full-project builds. The weights carry the MIT license. The Decoder reports that this permits commercial use, modification and redistribution.
The model has 1.6 trillion parameters in total. Meituan says about 48 billion are activated for each token. The Decoder reports that the hosted API can take in as much as one million tokens in a single request. It also reports that the model reads text, images and video.
How a model this large stays affordable to run
LongCat-2.5 uses a design called mixture of experts. The network is split into many specialised parts, and only a few are switched on for any given token. That is why 48 billion of the 1.6 trillion parameters do the work at any moment.
Meituan also compresses the model. It stores the experts' weights and runs their calculations in 4-bit numbers instead of 8-bit. Smaller numbers mean less data to move between chips. Meituan says the 4-bit format runs natively on domestic AI accelerators.
In Meituan's stress tests against an 8-bit baseline, prompt processing was 43.3% faster and response generation 42.77% faster. Time to first token fell 15.91%. Meituan says accuracy stays comparable to the uncompressed model on the benchmarks it evaluated. These are the company's own results.
The hardware path is the news
The Decoder reports that the model was trained entirely on Chinese hardware. Meituan's own card is narrower. It says the model's capabilities rest on infrastructure optimisation for domestic accelerators. The card does not name a chip maker.
Huawei appears in the self-hosting guide, according to The Decoder. It describes a cluster built from 194 Ascend accelerators in Huawei's Atlas A2 series. Two of them process images, 64 ingest prompts and 128 generate responses. The Decoder found no instructions for Nvidia GPUs.
Two lessons follow. First, "open weights" does not mean "easy to host". The guide's reference deployment is a 194-accelerator cluster, a data-centre project rather than a weekend install. Second, the supported hardware path is a product decision. If your infrastructure plan assumes one chip vendor, a model tuned for another gives you less to work with.
What the benchmarks do and do not show
Unmarked scores in Meituan's tables are measured in-house and, per the card, "indicative only". On Terminal Bench v2.1, which tests command-line tasks, Meituan reports 88.4 for LongCat-2.5. In the same table, Claude Opus 5 scores 89.1 and GPT-5.6 Sol 88.8. Deepseek-V4.1-Flash tops it at 90.6.
The picture changes on Terminal Bench v4.0, the harder variant, and on DeepSWE. On Terminal Bench v4.0, Meituan lists LongCat-2.5 at 38.5, against 37.3 for GPT-5.6 Sol and 53.9 for Opus 5. DeepSWE tests bug fixes in real code repositories. There LongCat-2.5 scores 64.6, and The Decoder says only Claude Opus 4.8 sits below it in the table. Deepseek-V4.1-Flash leads at 74.2, with GPT-5.6 Sol at 72.7.
The Decoder also notes what the table leaves out. The newer Claude Opus 5.5 and GPT-6.1 Sol are missing from it. On BrowseComp, a web-search test, Meituan reports 91.9, the highest score in its table. It ran that test in its own environment.
Taken together, this describes a model that is competitive on some agent tasks and behind on others. Independent testing on your own codebase would settle it.
Questions to put to your team
Start with the licence. The MIT terms are permissive, but the card says the licence grants no rights to Meituan trademarks or patents. Ask legal what that means for your use.
Ask where the model would run and who would operate it. If the answer is a hosted API, check the price after the promotion. The Decoder says Meituan's launch price is a limited-time offer of roughly €0.27 per million input tokens.
Ask for a test on your own tasks before trusting any vendor table. The card warns that the model has not been evaluated for every downstream use. It also advises assessing accuracy, safety and fairness before sensitive deployments.
Finally, ask how exposed your AI plans are to a single chip supplier. LongCat-2.5 does not answer that question. It shows that, on Meituan's numbers, a competitive agent model can arrive with a different hardware assumption than the one many buyers carry.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Meituan (LongCat team).
