It officially published LongCat-2.0 and confirmed that it was an anonymous model previously operated on OpenRouter under the name “Owl Alpha”. The model uses open source permits, with parameters equal to 1.6 trillions of dollars, and is located in proxy programming and generic reasoning tasks.
Run anonymously for two months
The model has accumulated calls on multiple platforms prior to its formal appearance. The Corps says it ranked first in the Hermes Agen workspace, second in Claude Code and third in OpenClaw deployment. The company also indicated that this was the first trillion-billion parameter model that was trained and deployed on the national ASIC.
Details of training and architecture
LongCat-2.0 uses a hybrid specialist structure, each token activates approximately 48 billion parameters, and complex issues can switch between 33 billion and 56 billion. The Mission claims that pre-training covered more than 35 trillion tokens, using more than 50,000 domestic accelerators, and that the training process did not result in a roll-back or an unrecoverable surge in losses.
Prices below multiple competitions
For API prices, the standard fee for LongCat-2.0 is $0.75 per million input token and $ 2.95 per million output. The price of the release promotion period was reduced to $0.30 and $1.20, with a creditable context read free of charge. It also launched a $1 billion token approximately $60 package for developers and heavy users.
Baseline test performance
On SWE-bench Pro, LongCat-2.0 scored 59.5, slightly higher than the GPT-5 58.6. In the FORTE test, it received 73.2 points, which is the same as Claude Opus 4.6, but lower than 77.8 GPT-5.5. The group has not yet published the opening hours of the model weights, and the GitHub and Hugging Face pages still show " On-line ".