Meituan Releases LongCat-2.5 Large Model, 1.6 Trillion Parameters Trained Entirely on Domestic Compute

Related topics top start Related topics top end Futures promotion begin Futures promotion end Miaopai begin Miaopai end Video player start Video player end Market chart begin Market chart end

Sina Tech News, November 2 morning news, Meituan released a new-generation large model LongCat-2.5, a native multimodal model oriented toward long-horizon agent tasks, with 1.6 trillion total parameters, and the entire training process was completed using domestic computing power.

In June this year, Meituan released LongCat-2.0, which for the first time completed the full-process training and inference of a trillion-parameter model on domestic computing power. According to the introduction, LongCat-2.5, on the basis of full-process domestic computing power training, further strengthens native multimodal capabilities, converts visual input into tokens that the model can process uniformly, improves visual understanding, and supports tasks such as interactive web pages and 3D display pages. In agent programming, the model adopts large-scale long-horizon reinforcement learning to improve end-to-end engineering task capabilities and better support complex software development tasks.

In the mid-training stage, LongCat-2.5 uses sequences of up to 1M tokens for long-context training, with cumulative training volume reaching trillions of tokens. When multiple computing cards collaboratively process such long sequences, memory capacity and inter-card communication directly affect efficiency. To this end, Meituan worked with domestic computing power vendors to jointly optimize memory usage, inter-card communication, and computing and transmission efficiency. In response to the challenge of small per-card memory in domestic computing power, Meituan optimized task division around supernodes, so that the main communication of parameters and task data is completed within the high-speed interconnect network inside the supernode.

At Meituan's second-quarter earnings call on August 28 this year, CEO Wang Xing said: The long-term value of AI capabilities lies in whether it can truly enter real workflows, solve real problems, and form long-term stable infrastructure capabilities. "We will not become a token factory. Our models and AI products will be used to support our main business and improve user and merchant experience and the company's operating efficiency."

Centered on the goal of "supporting the main business," LongCat-2.5 has focused on strengthening capabilities such as business analysis and software development, equipping local life merchants with an AI helper.

According to the LongCat-2.5 technical blog, in a case of multi-channel brand operation analysis, the model can autonomously build a channel profit statement, identify the growth quality problem that paid GMV (gross merchandise value) grew 23.2% while channel profit fell 1.7%, and, combining the economic value of clearance, user repurchase and inventory risk, propose differentiated resource allocation suggestions, completing a full analysis from business accounting and problem diagnosis to decision recommendations.

Meituan said that the release of LongCat-2.5 marks further progress for "domestic chips and domestic models." It not only natively supports multimodality, but the model's overall capabilities also rival current mature international closed-source models and leading open-source models. Model evaluation results show that in multiple mainstream benchmark evaluations, LongCat-2.5's scores fall between Claude-Opus-4.8 and Claude-Opus-5, and in some evaluations it surpasses Claude-Opus-5. news_keyword_pub,stock, iframe below the main text iframe below the main text General statement General statement QR code at the end start QR code at the end start

QR code at the end start QR code at the end start Editor name and work code

Editor name and work code