Meituan has revealed LongCat‑2.0, a 1.6‑trillion‑parameter open‑weight Mixture‑of‑Experts language model engineered for agentic coding workloads and natively supporting an ultra‑long 1‑million‑token context window. Positioned as Meituan’s next‑generation frontier‑scale coding and reasoning model, LongCat‑2.0 is designed to act as the core engine for autonomous software agents that must understand, generate, and execute code across real‑world repositories.
Frontier‑scale LongCat‑2.0 powers autonomous agents across real‑world repositories with native 1M‑token context
It targets workflows that combine code comprehension, tool use, and multi‑step decision making, aiming to deliver reliable long‑context reasoning over entire projects rather than isolated files. The release marks a significant evolution over earlier LongCat iterations, pairing a dramatic increase in total parameters with architectural refinements tuned for repository‑level analysis and complex task orchestration. By focusing on agentic coding rather than general chat, the model is positioned as infrastructure for future automated development environments and AI coding assistants. The model’s openly released weights, licensed under the MIT License, allow organizations to deploy LongCat‑2.0 across GPU and NPU clusters with minimal restriction.
At the architectural level, LongCat‑2.0 adopts a sparse Mixture‑of‑Experts design with 1.6 trillion total parameters, of which roughly 33 billion to 56 billion are dynamically activated per token through gated routing. This strategy keeps per‑token compute near an average of 48 billion parameters, allowing the model to scale capacity without incurring the cost of dense trillion‑parameter inference. The model’s architecture aligns with the principles of Mixture-of-Experts used in other large-scale AI systems.
Zero‑compute experts further reduce waste by skipping computation on simple tokens, while redirecting complex tokens to richer expert combinations. The model’s MOPD post‑training pipeline organizes experts into three groups dedicated to agent behavior, reasoning, and interaction, and a routing network selects among them according to task characteristics. This division is intended to balance tool‑using competence, multi‑hop analytical depth, and instruction following within a single unified system that can serve as both a coding engine and a general reasoning backend.
In addition, LongCat‑2.0 incorporates a large N‑gram Embedding module that expands capacity in a dimension orthogonal to the MoE experts, improving local language modeling and code tokenization, especially when traversing dense, repetitive source trees.
LongCat Sparse Attention underpins the model’s ability to operate over contexts up to one million tokens, maintaining throughput while avoiding the quadratic cost profiles of conventional attention. Instead of attending to every token uniformly, the mechanism prioritizes salient segments across layers and timesteps, enabling agents to retrieve relevant functions, configuration files, and documentation from an entire codebase in a single pass.
This long‑horizon optimization is integrated with training and deployment choices that preserve accuracy while keeping compute, memory, and throughput within the practical limits of Meituan’s domestic accelerator cluster. The result is a model that can follow extended chains of logic, track dependencies across thousands of files, and sustain coherent reasoning over hours‑long interactions without losing track of earlier steps.
Meituan trained LongCat‑2.0 from scratch on tens of trillions of multilingual and code tokens using a domestic ASIC supercomputing cluster with about 50,000 accelerator cards, and released the model weights openly to showcase large‑scale AI development on non‑Nvidia infrastructure to global developers.
