Meituan Launches LongCat-2.5: 1.6T-Param Multimodal Agent

Meituan on November 2, 2026 released LongCat-2.5 Preview, a 1.6-trillion-parameter mixture-of-experts model designed for long-horizon agentic work. With roughly 48 billion parameters active per token, the model targets sustained execution across terminals, browsers, spreadsheets, and design tools, backed by a 1-million-token context window and up to 128,000 tokens of output. The release is live through Meituan's chat interface and API, with GitHub and Hugging Face distributions planned.

Architecture and Scale

LongCat-2.5 uses a mixture-of-experts architecture with 1.6 trillion total parameters, activating approximately 48 billion per forward pass. The model supports a 1-million-token context window, allowing agents to hold entire codebases, multi-tab browser sessions, and multi-stage workflows in working memory. Maximum output is capped at 128,000 tokens. Meituan reports training on domestically produced Chinese AI accelerators, with W4A4 quantization applied for deployment efficiency.

The model is natively multimodal, processing visual and text inputs within a unified token framework rather than using separate encoders. This design supports what Meituan calls visual coding, creative design tasks, and 3D modeling workflows, with vision folded into the core architecture at the token level.

Training on Agent Failures

Meituan's training methodology leaned heavily on expanded long-task sequences and large-scale reinforcement learning for whole-codebase projects. The company says it converts failures encountered by real agents into reproducible training tasks, feeding breakage logs and error traces back into the curriculum. This closed-loop approach aims to build resilience into the model's decision-making for multi-step agent work, where cascading errors can derail long executions.

The training corpus prioritized sequences that require sustained reasoning and tool use, with an emphasis on coding, GUI automation, and document workflows. Vision data was integrated at the token level, allowing the model to reason over screenshots, diagrams, and design mockups alongside code and text.

Release Date and Availability

Officially launched on November 2, 2026, LongCat-2.5 is available now through Meituan's online chat service and API endpoints. The company has announced plans to release the model on GitHub and Hugging Face, though no specific timeline for those distributions has been provided. Access is currently managed through Meituan's platform, with no public pricing or rate-limit details disclosed at launch.

What This Means

LongCat-2.5 represents Meituan's push into long-context agentic AI, competing with models like DeepSeek V4 and other Chinese providers in the multimodal agent space. The 1-million-token context and 128,000-token output put it among the highest-capacity models for sustained agent work, while the training-on-failures approach suggests a focus on real-world robustness over benchmark performance. The domestic accelerator training reinforces China's hardware independence push. Meituan's emphasis on agent workflows positions LongCat-2.5 as infrastructure for automation in coding, design, and productivity tools, with the model's ability to process visual inputs at the token level opening use cases beyond text-only agents.