Meituan has now twice trained a trillion-scale model on domestic chips, and it is pointing that capacity at its merchants rather than selling tokens to developers.
Meituan released LongCat-2.5 on November 2, a native multimodal model with 1.6 trillion total parameters that the company says was trained entirely on domestic computing power. About 48 billion of those parameters are active for each token, TechNode reported, and the model is built for long-horizon agent work such as completing whole software projects inside large codebases. Caixin Global reported that Meituan plans to use it to raise efficiency across its core businesses.
The domestic-compute claim is the second in a row for the food delivery group. In June, Meituan said LongCat-2.0, a model of the same 1.6-trillion-parameter scale, was trained and served end to end on Chinese chips, as EastFrontier reported at the time. Meituan called the LongCat-2.5 run the largest training task yet completed on domestic compute, Leiphone (雷峰网) reported. Meituan has not named its chip supplier.
Performance is the open question. Meituan says LongCat-2.5 lands between Anthropic’s Claude Opus 4.8 and Opus 5 on its benchmarks, but every figure comes from the company’s own testing, and Tech Times noted that no independent evaluation had been published. The model is available through Meituan’s chat service and API platform, with GitHub and Hugging Face releases listed as coming soon.
LongCat-2.5 adds native vision to a model built for long agent tasks
The main technical change is how the model handles images. LongCat-2.5 converts visual signals into multi-level discrete tokens so that a single backbone processes pictures and text together, IT Home reported. Meituan says this supports visual coding, creative design and 3D tasks, such as building an interactive web page from an image.
On the agent side, Meituan extended training on long sequences of tasks and used large-scale reinforcement learning, a method that rewards the model for completing multi-step jobs, to improve its work on complex codebases. It also turns failures that real agents ran into back into reproducible training tasks, according to TechNode. During mid-training the model saw sequences of up to 1 million tokens, Xinhua reported, citing Meituan’s technical blog, and cumulative training ran to several trillion tokens, according to Huanqiu (环球网).
The full release follows a preview version that reached Meituan’s API platform on September 25 with image understanding and coding functions. Meituan gave the preview’s size then, about 1.6 trillion parameters with about 48 billion active in a mixture-of-experts design, IT Home reported at the time. The November 2 launch added the technical blog and the claim about domestic training.
Meituan says its gains come from tuning software to Chinese hardware
Most of the efficiency figures Meituan published concern how it got more out of domestic chips. In long-context fine-tuning tests, training throughput per card rose 86% compared with a standard distributed baseline, according to the technical blog cited by Xinhua. Meituan said it worked with domestic chipmakers on memory use, communication between cards and data transfer, and used high-speed links inside supernodes, the rack-scale systems that tie many accelerators together.
For serving the model, Meituan stores its expert weights and activations at 4-bit precision, a compression step known as W4A4 quantization that cuts the memory and compute each answer needs. The company says output quality matched the uncompressed model on the tasks it evaluated. Against an 8-bit (W8A8) baseline in online stress tests, throughput rose about 40%, the wait for the first token fell about 16%, and the time for each subsequent token fell about 20%, Xinhua reported.
The chip vendor remains unnamed. Tech Times wrote that the developer community has pointed to Huawei’s Ascend 910C, and pointed out that US Commerce Department guidance from May 2025 said using Ascend chips anywhere could breach US export controls. Meituan’s materials do not address the hardware supplier or that guidance.
Merchants are the first market, and open weights are the next marker
Meituan is aiming the model at its own commercial ecosystem. Caixin reported that LongCat-2.5 is meant to help merchants analyze channel profits, detect growth problems and propose how to allocate resources. In a brand analysis case, LongCat-2.5 built a channel profit statement and flagged that payment GMV had grown 23.2% while channel profit fell 1.7%. It then suggested shifting resources between channels, according to Xinhua.
The company has built an AI operations workbench called Miaoshou (妙手) on the LongCat series. It helps restaurant and hotel operators read industry metrics and brand rules and carry out tasks through Meituan’s tools within the permissions merchants grant it. Chief executive Wang Xing set out the logic on the company’s August earnings call: "We will not become a token factory; our models and AI products will support our core business," he said, as quoted by Xinhua.
The next test is outside scrutiny. LongCat-2.0’s weights were published openly, which let developers measure the model themselves. If Meituan posts LongCat-2.5 to Hugging Face as planned, third-party benchmark runs will show whether the Opus-range claims hold, and the accompanying technical documentation may finally identify the chips that trained it.