Meituan releases LongCat-2.5, an AI model aimed at whole software projects

Loading page…

Meituan released LongCat-2.5 on November 2, 2026, putting a model designed for extended AI coding tasks into its chat service and developer API. The company’s pitch reaches beyond generating individual pieces of code: it says the model combines visual understanding, tool use and multi-step reasoning to work on complex codebases and complete software projects.

From preview to longer assignments

The release follows LongCat-2.5-Preview, listed in the official API changelog on September 25. That version already offered image understanding, coding and compatibility with development tools.

Meituan says it expanded training on long sequences of tasks and used large-scale reinforcement learning to improve work on complex code and complete projects, according to TechNode. It also turns failures encountered by real agents into reproducible training tasks. That gives the training effort a specific source of problems to address: work that agents have already struggled to complete, rather than coding exercises alone.

Its release notes describe intended capabilities, however, without publishing task-completion scores that quantify how reliably the model finishes a whole project.

Total size versus active computation

1.6 trillionTotal parameters

TechNode reports that LongCat-2.5 has 1.6 trillion parameters in total.

Approximately 48 billionActivated per token

Approximately 48 billion parameters are activated for each token, a separate measure from the model’s total size.

Visual inputs meet coding tools

LongCat-2.5 processes visual and text inputs within a unified token framework, TechNode reports. The official description connects that visual understanding directly to coding, creative design and 3D tasks. The model is therefore being positioned for work involving both images and code, not only text instructions followed by a written response.

The release also targets developers who already use software to run coding agents. LongCat lists compatibility with Claude Code, Codex, Hermes and other agent harnesses—the surrounding tools used to run an agent’s work.