Meituan releases LongCat-2.0, a 1.6T parameter MoE model optimized for agentic coding.
LongCat-2.0's 48B active parameters within a 1.6T MoE architecture places it in the heavyweight class of coding models, directly rivaling top-tier Western counterparts. Meituan's focus on agentic workflows trained on a massive 30T token corpus signals an aggressive push into specialized software engineering automation. If inference efficiency matches the architectural scale, this will be a highly disruptive tool for autonomous dev environments.
Meituan has unexpectedly entered the heavyweight AI race with the release of LongCat-2.0, a massive Mixture-of-Experts (MoE) model explicitly engineered for agentic coding tasks. While Meituan is traditionally known for its e-commerce and delivery platforms, this release positions them as a formidable player in the foundational model space.
Technical Breakdown LongCat-2.0 features a staggering 1.6 trillion total parameters, utilizing a sparse MoE architecture that activates approximately 48 billion parameters per token during inference. This places it in the same architectural weight class as top-tier proprietary models like GPT-4 or Grok-1.5. Most notably, the model was trained on an exhaustive corpus of 30 trillion tokens. For a model specialized in agentic coding, a dataset of this magnitude implies aggressive scraping and curation of code repositories, execution logs, issue trackers, and multi-step tool-use trajectories.
Why It Matters From an engineering perspective, the shift from standard code completion to "agentic coding" is the current frontier. Agentic models must excel at multi-step reasoning, environment interaction, and self-correction. By activating 48B parameters per token, LongCat-2.0 strikes a deliberate balance between immense representational capacity and inference efficiency. If the routing mechanism in its MoE layers is highly optimized, it could offer a powerful engine for autonomous software engineering workflows, rivaling specialized tools like Devin or SWE-agent. Furthermore, this release underscores China's rapidly maturing AI ecosystem, proving that companies beyond the usual tier-one AI labs possess the compute and data pipelines to train frontier-class models.
What to Watch Next The immediate metric to watch will be LongCat-2.0's performance on standard agentic benchmarks like SWE-bench, as well as its context window capabilities—which are crucial for repository-level reasoning. Engineers should also look for details on its deployment footprint: serving a 1.6T parameter model, even with sparse activation, requires significant VRAM across multi-node clusters. Finally, whether Meituan releases the model weights openly or keeps it gated behind an API will determine its adoption rate within the broader developer community.