❯ Meituan releases LongCat-2.5 with stronger vision and long-running agent capabilities

Built for tasks with many stepsMeituan released the natively multimodal LongCat-2.5 on November 2. The 1.6-trillion-parameter model was trained entirely on domestic computing hardware. According to Meituan’s announcement, it builds on June’s LongCat-2.0 with stronger visual understanding and agentic coding. It is accessible on the LongCat website and through an API.

Connect perception with executionNative multimodality lets the model process text and visual inputs together for interactive websites, 3D presentation pages and complex software development. Long-running tasks require repeated tool calls, observations and revisions. Meituan uses large-scale reinforcement learning on extended tasks to improve that execution process.

Reduce the computing burdenTraining included sequences of up to one million tokens, the units used to process text and visual information. Meituan optimized memory use and communication between domestic accelerators and adopted W4A4 inference, representing expert weights and activations in four bits. The company reports roughly 40% higher input and generation throughput than its W8A8 baseline in online load tests.

Apply the model to merchant operationsMeituan uses the LongCat family in its Miaoshou AI workspace for channel-profit analysis, store management and customer acquisition. Combining vision, coding and tool use allows the system to organize business data, diagnose problems and act through merchant-authorized systems.

▮ SIGNALLong-running agents need better execution and computing efficiency together to become practical tools for everyday business operations.