Model WarsJuly 5, 2026via MarkTechPost
Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention
Why it matters
Meituan's LongCat-2.0 signals China's AI infrastructure independence and competitive parity in large-scale open models. The 1M native context window and sparse attention architecture represent a meaningful capability milestone, but vendor benchmarks require independent verification.
Key signals
- 1.6 trillion total parameters, ~48B active per token (MoE)
- Native 1M token context window via LongCat Sparse Attention
- Trained and served on domestic AI ASIC superpods (no NVIDIA dependency)
- Open model release (availability/licensing not fully detailed)
- Vendor-reported benchmarks—independent validation pending
- Suggests China's AI compute self-sufficiency at scale
The hook
1.6T parameters. 1M context. Meituan just open-sourced a MoE model that rivals frontier labs—trained entirely on domestic AI ASICs.
Meituan has released LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts model that activates about 48 billion parameters per token. It pairs a native 1-million-token context, built on LongCat Sparse Attention, with training and serving run end-to-end on domestic AI ASIC superpods. Here is the architecture, the vendor-reported benchmarks, the API access path, and what remains unverified.