Meituan has launched LongCat-2.5, a native multimodal AI model aimed at extended coding and “long-horizon” agentic tasks. The release, dated November 2, adds improved visual understanding alongside coding capabilities, and positions the model for work that requires sustained execution over long sequences rather than short responses.
Across the coverage, both outlets describe the model’s scale and design: LongCat-2.5 has 1.6 trillion parameters in total, with roughly 48 billion parameters activated per token. Both also link performance to training approaches that support long task execution, including reinforcement learning geared toward agentic coding. TechNode additionally frames the launch as an expansion of Meituan’s AI model lineup and notes training on long task sequences. Pandaily emphasizes optimization for inference on domestic compute and highlights W4A4 inference alongside the long-horizon agentic coding focus.