LongCat-2.0: Meituan releases 1.6T-parameter MoE model with 48B active and 1M context

What happened

Meituan's LongCat team published the LongCat-2.0 model card and blog post on June 30, with the release climbing into the top of Hacker News.

Context and impact

LongCat-2.0 is another data point in the rapid rise of Chinese open-weight models that match or approach Western frontier systems while avoiding US-controlled accelerators. It strengthens the argument that China can build trillion-parameter systems entirely on domestic silicon despite export controls, and gives developers worldwide another large, permissively-available reasoning model. The release is also notable because Meituan, a food-delivery and services giant, is now firmly inside the frontier-lab race.

Details

• 1.6 trillion total parameters, ~48B active per token
• 1M-token context, optimized for agentic scenarios including coding and task planning
• Trained on 50,000–60,000 domestic Chinese chips — the largest domestic-hardware training run in China
• Each user gets 10M tokens per day free during the test period
• Parametrically comparable to the recently released DeepSeek V4