Meituan open-sourced LongCat-2.0 on June 30, a 1.6-trillion-parameter Mixture-of-Experts model built for agentic coding. It activates only about 48B parameters per token, dynamically swinging between 33B and 56B depending on how hard the query is, and carries a native 1M-token context window. The GitHub repo ships it under an MIT license.
Meituan reports 59.5 on SWE-bench Pro, edging out GPT-5.5's 58.6, plus 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. All self-reported, and independent testing hasn't confirmed them. The repo's own comparison table still puts LongCat behind Claude Opus 4.8 on broader agent benchmarks like FORTE and BrowseComp, so the win is narrow and specific to software engineering.
The architecture leans on four tricks: LongCat Sparse Attention to drop long-context cost toward linear, zero-computation experts that route trivial tokens like punctuation through unchanged, a shortcut-connected ScMoE backbone, and MOPD post-training that splits work across agent, reasoning, and interaction expert groups. Pretraining ran from scratch on 35T+ tokens.
The part drawing the most attention isn't the benchmarks. Meituan says the entire training and inference cycle ran on a 50,000-card cluster of domestic Chinese ASICs, no Nvidia hardware involved. The model had already been running anonymously as "Owl Alpha" on OpenRouter for roughly two months before the reveal. One catch worth flagging: several outlets noted the full weights were marked "coming soon" at launch, so open-source in practice may have lagged the announcement. It deploys on both GPU and NPU.
Bottom Line
LongCat-2.0 scores 59.5 on SWE-bench Pro, just past GPT-5.5's 58.6, and was trained entirely on 50,000 domestic Chinese ASICs.
Quick Facts
- 1.6T total parameters, ~48B active per token (range 33B-56B)
- Released June 30, 2026 under MIT license
- SWE-bench Pro 59.5, Terminal-Bench 2.1 70.8, SWE-bench Multilingual 77.3 (all company-reported)
- Native 1M-token context window
- Pretrained from scratch on 35T+ tokens; trained on 50,000-card domestic ASIC cluster
