Meituan Open-Sources LongCat-2.0: 1.6T MoE Trained on Domestic Chips

赛博禅心 · wechat · 2026-07-05

Meituan has officially open-sourced the large model LongCat-2.0, a 1.6 trillion parameter MoE architecture with dynamic activation between 33B and 56B, natively supporting a 1 million token context. Across six official Agent benchmarks, it ranks just behind Claude and GPT, performing on par with Gemini 3.1 Pro.

The biggest highlight is that training and inference run entirely on domestic AI ASIC clusters: peak training utilized over 50,000 domestic cards with 0 rollbacks, keeping VRAM under 60GB (matching the 60GB HBM of domestic cards). To alleviate cross-card communication bottlenecks, the team adopted the ScMoE architecture to parallelize Dense FFN computation with MoE communication, theoretically halving inference latency. They also introduced an "empty expert" mechanism for dynamic activation—simple tokens activate 33B, while complex tokens max out at 56B.

The million-token context uses the in-house LongCatSparseAttention (comparable to DeepSeek V3.2's DSA), reducing scoring overhead by reusing indices across adjacent layers and applying a coarse-to-fine block filtering approach. Additionally, custom high-performance deterministic operators pushed steady-state daily training throughput to 1T tokens. A preview version anonymously hit OpenRouter as OwlAlpha for two months, ranking top three globally in monthly invocations, and supports switching via Codex.

Related event: Meituan Open-Sources LongCat-2.0 on Domestic Chips(3 posts)→

Original post →

More from Infra

Infra channel →