Meituan Open-Sources 1.6T Parameter Multimodal Model LongCat-2.5 with 1M Context

On November 1, 2026, Meituan's LongCat team unveiled LongCat-2.5, a native multimodal model designed for long-horizon agentic tasks. One week later, on November 9, the company officially released the model's weights under the MIT license, permitting commercial inference deployment and secondary training without additional authorization. LongCat-2.5 is a 1.6T-parameter MoE model, with approximately 48B parameters activated per token, and supports a context window of 1 million tokens.

The model's visual capabilities stem from a fundamental redesign of its representation layer. Rather than relying on an external vision module, the model encodes visual signals as multi-level discrete tokens that are processed uniformly by the backbone alongside text. Specifically, it applies residual vector quantization to native-resolution visual features to progressively recover detail; visual tokens can be precomputed and cached offline, reducing online compute costs and balancing distributed training loads; and the visual codebook is jointly trained with the backbone, allowing vision and text to share the same parameters and semantic space. This architecture notably enhances the model's performance in Visual Coding, creative design, and 3D tasks. Additionally, the team performed deep co-optimization across model architecture, chip adaptation, and deployment strategies to accommodate domestic compute chips with limited memory and bandwidth.

For training methodology, Meituan expanded the entire pipeline rather than making isolated changes. The team incorporated more long-horizon trajectories and converted real-world Agent failure traces into reproducible, verifiable training tasks. They adopted partial rollout asynchronous training to save and resume unfinished long tasks, dynamically adjusted oversampling rates based on pass rates, and applied gradient correction to stale trajectories spanning multiple checkpoints. To discourage reward hacking, the training environment disconnects from the network and clears git history and file caches, while an LLM judge supervises traces. The team also imposed constraints on repetitive thinking, duplicate tool calls, and overthinking to keep the model focused on task completion. Meituan states that LongCat-2.5 delivers improved end-to-end capabilities in complex codebase development, complete project construction, Visual Coding, and professional office scenarios.

Following the open-source release, the open-source AI coding tool OpenCode immediately announced a two-week free trial for LongCat-2.5, supporting multimodal inputs and 1M token contexts with a daily limit of 100 requests. According to OpenCode's public data, by 12:00 PM on November 10, the model had accumulated approximately 109B tokens, attracted around 4,700 unique users, and completed about 50,000 sessions. In Meituan's published benchmark suite, LongCat-2.5 was compared against nine other models including Claude Opus 5, GPT-5.6-Sol, GLM-5.3, Kimi K3, and DeepSeek-V4.1-Flash, spanning 31 tests. The model weights are now publicly available for download.