LongCat-2.5

LongCat-2.5 is Meituan's latest native multimodal large model for long-horizon agentic tasks, released on November 2, 2026. It is a Mixture-of-Experts system with 1.6 trillion total parameters and about 48 billion activated per token. Compared with LongCat-2.0 (June 2026), it strengthens visual understanding and agentic coding, and keeps the "domestic-chip, domestic-model" training path on Chinese accelerators.

Meituan says the model can handle visual coding, creative design, 3D work, complex codebase changes and end-to-end project builds, and that it plugs into mainstream coding harnesses such as Claude Code, Codex and Hermes. Users can try it on longcat.ai; developers call it through the LongCat API. GitHub and Hugging Face weights are promised to follow.

What changed vs LongCat-2.0

• Native multimodality — visual signals become multi-level discrete tokens so the same backbone handles images and text.
• Long-horizon RL — more long trajectories, failure-to-task conversion, and mixed training across harnesses for multi-step engineering work.
• Domestic infra — long-context training up to ~1M tokens and W4A4 inference optimized on Chinese AI accelerators.

Positioning

Meituan frames LongCat as infrastructure for its local-commerce business (merchant ops, software assistance), not a pure "token factory". Early coverage places scores between Claude Opus 4.8 and Opus 5 on several public suites, with some wins over Opus 5.

Availability

Online playground + API as of Nov 2, 2026; open weights later.

1. Open https://longcat.ai for the chat playground, or create a LongCat API account.
2. Point an OpenAI-compatible client / Claude Code / Codex / Hermes at the LongCat endpoint and select LongCat-2.5.
3. Prefer long, tool-using tasks (multi-file refactors, visual-to-code, project scaffolding) where agentic coding matters.
4. Watch for the promised Hugging Face / GitHub release if you need local weights.