On November 2nd, Meituan released the new generation of large model LongCat-2.5, which is a native multimodal model for long-range intelligent agent tasks. The total parameters are 1.6 trillion, and the entire training process is completed by domestic computing power.

In June of this year, Meituan released LongCat-2.0, which completed the full process training and inference of a trillion parameter model on domestic computing power for the first time. On the basis of LongCat-2.0, the new model further enhances native visual understanding and intelligent agent programming capabilities, and optimizes the efficiency of long context training and inference on domestic computing power.

The release of LongCat-2.5 marks a further step forward for "Guoxin Guomo": not only does it natively support multimodality, but its overall model capability is also comparable to current international mature closed source models and top open source models.

▲ LongCat-2.5 scores in multiple mainstream benchmark evaluations are between Claude-Opus-4.8 and Claude-Opus-5 (see data chart)

The model evaluation results show that in multiple mainstream benchmark evaluations, LongCat-2.5 scores between Claude-Opus-4.8 and Claude-Opus-5, and has surpassed Claude-Opus-5 in some evaluations.

Users can visit longcat on the official websiteAI experience, developers can call models through synchronized online APIs.

Enhance native multimodal capabilities and improve domestic computing power training and inference efficiency

On the basis of domestic computing power training throughout the entire process, LongCat-2.5 further strengthens its native multimodal capabilities, converts visual input into tokens that can be uniformly processed by the model, enhances visual understanding, and supports tasks such as interactive web pages and 3D display pages. In terms of intelligent agent programming, the model adopts large-scale long-range reinforcement learning to enhance end-to-end engineering task capabilities and better support complex software development tasks.

LongCat-2.5 uses sequences of up to 1M tokens for long context training during the mid training phase, with a cumulative training volume of trillions of tokens. When multiple computing cards collaborate to process such long sequences, the video memory capacity and inter card communication directly affect efficiency. To this end, Meituan collaborates with domestic computing power manufacturers to optimize the use of video memory, inter card communication, and computing and transmission efficiency. In view of the small challenge of the display memory of the domestic computing power single card, Meituan optimizes the task division around the super node, so that the main communication of parameters and task data is completed in the high-speed Internet within the super node.

According to the technical blog of LongCat-2.5, in the measured long context supervised fine-tuning task, the optimized single card training throughput increased by 86% compared to the DDP (Distributed Data Parallel) baseline. According to the throughput of a single card, it is equivalent to one card being able to complete the workload of about 1.86 cards in the same time, and the processing efficiency is nearly doubled.

In terms of reasoning, LongCat-2.5 adopts the W4A4 scheme, which quantifies expert weights and activations to 4 bits simultaneously, representing the data required for computation with a lower bit width, and optimizes computation and transmission based on the hardware characteristics of domestic AI accelerators. On the downstream tasks that have been evaluated, this approach maintains comparable performance to the model before quantification.

Online pressure testing shows that compared to the W8A8 INT8 baseline with 8-bit weights and activations, the throughput of input processing and token generation has increased by about 40%, which means that more inputs can be processed and more content can be generated in the same time. The waiting time for the first output is reduced by about 16%, and the subsequent generation time for each token is reduced by about 20%. The time for users to wait for the model output is also correspondingly reduced.

From native multimodal capabilities to training and inference efficiency, LongCat-2.5 continues the development path of the LongCat series' National Chip and National Model 'and further expands the task boundaries that the model can handle.

Expand core business capabilities and equip merchants with AI assistants

At the Meituan Q2 earnings conference call on August 28th this year, CEO Wang Xing stated that the long-term value of AI capabilities lies in whether they 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, enhance user and merchant experience, and improve company operational efficiency. ”

Centered around the goal of "supporting the main business", LongCat-2.5 has focused on strengthening its capabilities in business analysis, software development, and equipping local lifestyle businesses with AI assistants.

According to the LongCat-2.5 technical blog, in the case of brand multi-channel management analysis, the model can independently construct a channel profit statement, identify the growth quality problem of a 23.2% increase in payment GMV (total merchandise transaction volume) and a 1.7% decrease in channel profit, and propose differentiated resource allocation suggestions based on the economic value of clearance, user repurchase, and inventory risk, completing a complete analysis from business accounting, problem diagnosis to decision-making recommendations.

Based on the LongCat series model, Meituan has launched the AI management workbench "Miaoshou" to equip local merchants with AI assistants. (Data image)

Based on the LongCat series model, Meituan has launched the AI management workbench "Miaoshou" to help merchants understand industry indicators and brand rules in high-value business scenarios, and call tools and systems within the authorized scope of merchants to complete business task execution.

At present, "Miaoshou" has served chain brands such as Guming, Xushengji, and Yihetang, integrating high-value business scenarios such as brand management analysis, store management, and customer growth, helping businesses move from problem discovery to decision-making and execution.