Meituan Launches LongCat-2.5: 1.6T-Parameter Native Multimodal Model for Long-Horizon Agent Tasks

What happened

The new model represents visual signals as multi-level discrete tokens, so the backbone handles vision and text in a single token form, supporting interactive web pages and 3D showcase pages. The official introduction stresses two upgrades: native visual understanding combined with agentic coding, and end-to-end engineering ability on long-horizon tasks, covering complex codebase development and full project builds.

On training, it extends the LongCat-2.0 route: LongCat-2.0, released in June 2026, was the first trillion-parameter model trained and run end to end on domestic compute, and LongCat-2.5 was likewise trained entirely on domestic compute. Long-sequence training added more long-trajectory data, and large-scale long-horizon reinforcement learning raised capability; mid-training used sequences up to 1M tokens, with cumulative training volume in the trillions of tokens.

Inference is likewise tuned for domestic accelerators: a W4A4 scheme quantizes expert weights and activations to 4 bits, with train-inference numerical consistency and long-context rollout optimization supporting large-scale long-horizon reinforcement learning. The technical blog says optimized single-card training throughput on the tested long-context supervised fine-tuning task rose 86% over a distributed data parallel baseline. Users can try it at LongCat.ai, developers can call the newly launched API, and GitHub and Hugging Face versions follow later.

Key facts

1. Publisher and date: Meituan's LongCat team, Nov 2, 2026
2. Scale: 1.6 trillion total parameters, about 48 billion activated per token
3. Positioning: a native multimodal model for long-horizon agent tasks, strengthening native visual understanding and agentic coding
4. Training: entirely on domestic compute; more long-trajectory data in long-sequence training plus large-scale long-horizon reinforcement learning; mid-training sequences up to 1M tokens, cumulative volume in the trillions of tokens
5. Inference: W4A4 scheme quantizing expert weights and activations to 4 bits, tuned for domestic AI accelerators
6. Benchmarks: the company says scores on several mainstream benchmarks fall between Claude Opus 4.8 and Claude Opus 5, and above Opus 5 on some; the figures come from the official release, not an independent evaluation
7. Availability: trial at LongCat.ai, API launched alongside, GitHub and Hugging Face versions to follow