sophnet/LongCat-Flash-Chat

LongCat-Flash-Chat is an efficient Mixture-of-Experts (MoE) language model officially released and open-sourced by Meituan’s tech team. Its core positioning is as a next-generation AI foundational model characterized by “dynamic computation, ultra-fast inference, and agent-first architecture,” specifically designed for complex, long-duration, and tool-intensive tasks.

• Ultra-large-scale MoE architecture: With a total parameter count reaching 560B, each token activates only between 18.6B and 31.3B parameters (averaging around 27B). This is achieved through an innovative “Zero-Computation Experts” mechanism that enables on-demand allocation of computational resources.
• Industry-leading inference efficiency: On NVIDIA H800, it achieves a generation speed of over 100 tokens/s, significantly faster than mainstream models of similar or even smaller scale.
• Comprehensively superior agent capabilities: It excels in the τ2-Bench (tool usage) and VitaBench (complex-scenario agents), making it particularly well-suited for multi-step agent applications that demand longer processing times.

───────────────────────────────────────────────────────────────────

Core Capabilities

🧠 Strong instruction-following ability: Ranked first in IFEval (89.65), and also achieved top scores in the Chinese instruction benchmarks COLLIE and Meeseeks-zh.

📚 Rigorous general knowledge: Scored 89.71 on MMLU, 90.44 on CEval, and 86.50 on ArenaHard-V2—its overall performance rivals that of China’s top models.

🧩 Natively trained for agent-based tasks: Built its own agent-specific evaluation dataset and adopted a multi-agent approach to generate high-quality trajectory data, optimizing the entire process from tool calls to environmental interactions.

⚡ System-level engineering optimization: Introduced technologies such as cross-layer communication parallelism and customized low-level operators, enabling efficient training within just 30 days and achieving extremely low inference latency.