LongCat-2.0
1.6T-total MoE with 1M context, trained on 35T+ tokens entirely on Chinese AI accelerators; 59.5 on SWE-bench Pro, weights under MIT.
LongCat Sparse Attention improves on DeepSeek's DSA indexer; 135B N-gram embedding parameters sit alongside MoE. Pre-release ran anonymously on OpenRouter as 'Owl Alpha' at about 559B tokens/day. Weights followed on 2026-07-05; Meituan says it used a Huawei communication library.
- Date
- Monday, 29 June 2026
- Lab
- Meituan (LongCat)
- Kind
- open-weights
- Access
- open weights
- Price
- $0.30 input / $1.20 output per M tokens promo; $0.75 / $2.95 standard (VentureBeat)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Pro | 59.5 vs GPT-5.5 58.6, Claude Opus 4.8 69.2 | company |
| Terminal-Bench 2.1 | 70.8 | company |
| BrowseComp | 79.9 | company |
| Pretraining tokens | 35T+ no rollbacks reported | company |
| Accelerators used | 50,000+ Chinese ASICs per VentureBeat | company |
Launch 2026-06-29 (VentureBeat said weights were still 'coming soon'); HF repo public from 2026-07-05. HF says ~48B active parameters, while VentureBeat says 33-56B (dynamic). Meituan did not name the chip vendor. Benchmarks are in-house.
Sources
- venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic
- huggingface.co/meituan-longcat/LongCat-2.0
- en.wikipedia.org/wiki/Meituan
This record was checked against its sources on 6 October 2026. How we check