Frontier Models
Meituan LongCat-2.0 1.6T MoE Model Edges GPT-5.5 on SWE-bench Pro
The open-sourced release from the Chinese delivery firm features 1 million token context and ASIC-based training while powering agent capabilities on OpenRouter.
LongCat-2.0 is a 1.6 trillion parameter Mixture-of-Experts language model from Meituan designed specifically for agentic coding tasks.
Meituan released LongCat-2.0 as an open source model.
The model targets agentic coding applications.
What are the core specifications of LongCat-2.0?
LongCat-2.0 maintains 1.6 trillion total parameters.
Approximately 48 billion parameters activate per token.
The model supports a native 1 million token context window.
LongCat-2.0 incorporates LongCat Sparse Attention.
N-gram Embedding modules form part of the architecture.
How was LongCat-2.0 trained?
Both the full training run and large-scale deployment rely on AI ASIC superpods.
Pretraining spans millions of accelerator-days.
The process covers more than 35 trillion tokens.
Training involved more than 50K AI ASICs.
No rollbacks or irrecoverable loss spikes occurred.
What benchmarks does LongCat-2.0 achieve?
LongCat-2.0 scores 59.5 on SWE-bench Pro.
GPT-5.5 scores 58.6 on the same benchmark.
LongCat-2.0 scores 70.8 on Terminal-Bench 2.1.
We are introducing and open sourcing LongCat-2.0, a large-scale MoE language model with 1.6 trillion total parameters and ~48 billion activated per token — a substantial step up from previous LongCat models, accompanied by several architectural improvements.LongCat team
What specialized components support agentic coding?
Specialized experts form distinct groups.
One expert group handles Agent tasks.
A separate group manages Reasoning.
Another expert group addresses Interaction.
How does LongCat-2.0 connect to platforms?
LongCat-2.0 serves as the model behind Owl Alpha on OpenRouter.
Open sourcing enables community access to the full model.
| Specification | Value |
|---|---|
| Total Parameters | 1.6 trillion |
| Active Parameters | ~48 billion |
| Context Window | 1 million tokens |
| Pretraining Tokens | >35 trillion |
| Training Hardware | AI ASIC superpods |
| Key Benchmarks | SWE-bench Pro: 59.5, Terminal-Bench 2.1: 70.8 |
What steps allow users to engage with LongCat-2.0?
- Access the model weights from the Hugging Face repository.
- Review the full announcement and technical details on the LongCat blog.
- Evaluate performance through SWE-bench Pro and Terminal-Bench 2.1.
- Integrate the model via OpenRouter for Owl Alpha agent use.
- Experiment with LongCat Sparse Attention in custom agent workflows.
What implications follow for stakeholders?
Training on domestic AI ASIC hardware demonstrates scalable infrastructure.
Open sourcing at this scale adds to available frontier models.
Agentic coding optimization targets specific enterprise and developer needs.
What developments could come next?
Additional expert specializations may expand capabilities.
Further refinements to sparse attention could improve efficiency.
New fine-tunes or tools may build on the released base model.
Frequently asked
What benchmark scores does LongCat-2.0 report?
LongCat-2.0 achieves 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1. These figures exceed the GPT-5.5 result of 58.6 on SWE-bench Pro.