Meituan releases LongCat-2.5, a 1.6 trillion parameter agentic model

Chinese tech company Meituan has released LongCat-2.5, a native multimodal model built for long horizon agentic work such as full codebase builds. The model has 1.6 trillion total parameters with about 48 billion activated per token, and builds on the earlier LongCat-2.0 with stronger visual understanding and agentic coding ability.

Meituan trained the model on more long horizon trajectories and used large scale reinforcement learning aimed at complex, full project software development, converting real agent failure cases into reproducible training tasks. To keep the model from gaming its training rewards, Meituan disabled network access in task environments, removed git history and used an LLM judge to audit the resulting trajectories. Visual input is represented as discrete tokens alongside text, which the company says supports visual coding, creative design and 3D tasks.

On the infrastructure side, Meituan trained and runs inference using 4-bit weights and activations on domestic Chinese AI accelerators, each needing under 80 GB of memory, and reports throughput and latency improvements over an 8-bit baseline. In Meituan's own benchmark comparisons against models including GPT-5.5, Claude Opus 5 and DeepSeek-V4.1-Flash, LongCat-2.5 leads on some tests and trails on others. These figures are self-reported by Meituan and have not been independently verified. The model is available through Meituan's LongCat API and longcat.chat; no open weights have been released yet.

Why it matters

It shows a major Chinese lab pushing both model scale and domestic chip efficiency for agentic coding, without yet opening the weights for outside scrutiny.

Source: Meituan LongCat