LongCat-2.0 is a large language model released and open-sourced by Chinese local-services company Meituan on June 30, 2026 under the MIT license. It is a 1.6-trillion-parameter mixture-of-experts model with roughly 48 billion parameters active per token and a one-million-token context window. Meituan says it is China's biggest model trained entirely on domestic chips (Source: scmp.com; venturebeat.com).
Capabilities and reception
Coverage describes LongCat-2.0 as a near-frontier agentic coding model, and it has been leading OpenRouter usage rankings (Source: venturebeat.com).
Training and architecture
The model uses a mixture-of-experts architecture: 1.6 trillion total parameters with approximately 48 billion active per token, and a one-million-token context window. The claim that training ran entirely on domestic Chinese chips is the company's own; the specific chips used were not identified in initial coverage (Source: scmp.com). Reviewing the model on August 2, 2026, Florian Brand and Nathan Lambert identified the accelerators as Huawei Ascend 910s and described LongCat-2.0 as the first non-Huawei, non-toy model trained entirely on Chinese accelerators, noting that other Chinese chips are mostly used for inference. They rated the model as not the most capable for its size beyond benchmarks (Source: interconnects.ai). VentureBeat reported the training cluster at roughly 50,000 Chinese-made chips, described as an industry first at that scale, and that the model had quietly served production traffic on OpenRouter under the alias "Owl Alpha" before release (Source: venturebeat.com). If accurate, the release adds a data point to the debate over whether US chip export controls constrain Chinese frontier training (see Export Controls (AI), Semiconductor Supply Chain).
Availability
The model is open-source under the MIT license (Source: scmp.com).
Related models
LongCat-2.0 joins the cohort of Chinese open-weight models competing near the frontier, alongside DeepSeek V4, Qwen3, Kimi K2, and GLM-5.
Relationships
- instance-of: Open-Weight Frontier Models — MIT-licensed open release
- related: Export Controls (AI) — domestic-chip training claim bears on control efficacy
- related: US-China AI Competition: Different Races, Different Metrics — Chinese open-weight frontier cohort