Chinese food delivery giant Meituan has unleashed LongCat-Flash-Thinking, a 560-billion-parameter open-source AI model, on September 23, 2025. Evolving from LongCat-Flash-Chat, this Mixture-of-Experts model uses curriculum learning and large-scale reinforcement learning (DORA) to tackle complex math, coding, theorem proving, and agentic tasks. Activating ~27B parameters per token, it rivals OpenAI’s GPT-5, scoring 99.2% on MATH500, 81.6% on MiniF2F, and 79.4% on LiveCodeBench, while slashing token usage by 64.5% on AIME-25. With 93.7% accuracy in harmful content detection, it’s safe and sharp.
🧠 Freely Accessible Powerhouse
Released under the MIT license on Hugging Face and GitHub, LongCat-Flash-Thinking offers a free API with 500,000 daily tokens (expandable to 5M with approval). It generates 100+ tokens/second on H800 GPUs at ~$0.69/million tokens, making it a cost-effective beast. Users can interact at longcat.ai, with a “Think” mode for deep reasoning.
📍Who Gets It and When?
Available globally, the model’s weights and API are open to all, with higher quotas for approved users. Meituan’s internal tests show 20% efficiency gains in customer service, hinting at broader applications.
🧪Redefining AI Accessibility
Launched after rigorous training, LongCat-Flash-Thinking positions Meituan as a leader in China’s AI race, offering a transparent, powerful alternative to proprietary models. It’s not just code or math—it’s AI that thinks, solves, and scales, one query at a time.
