Meituan, the Chinese delivery giant, has officially launched LongCat-2.5, a native multimodal model with 1.6 trillion total parameters (about 48 billion active per token). The model is optimized for long-horizon agent tasks and was trained on domestic Chinese chips.
What is LongCat-2.5
Announced on November 1 and 2, 2026, LongCat-2.5 is a Mixture-of-Experts model focused on agentic coding and native visual understanding. It converts images into multi-level discrete tokens and processes text and vision in the same stream. It supports a 1 million token context and up to 128k token output.
According to the LongCat team, the model was trained with long agent trajectories and large-scale reinforcement learning to complete entire projects, not just answer questions. It is available via Meituan’s API at competitive pricing (around $0.30 input / $1.20 output per million tokens).
Performance and context
In benchmarks reported by Meituan itself, the model scored 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, competing with top models such as Claude Opus 5 and GPT-5.6 on several agent tasks. The launch highlights Chinese AI progress in agentic models trained locally.
Main source: Meituan Tech Blog.
Image credit: Meituan / reproduction
By GeekikiBot
