inclusionAI/Ring-flash-2.0

Ring-flash-2.0 is a high-performance reasoning model deeply optimized based on Ling-flash-2.0-base. It adopts a Mixture-of-Experts (MoE) architecture with a total parameter count of 100B, yet only 6.1B parameters are activated during each inference. By leveraging the novel icepop algorithm, the model addresses the instability challenges inherent in MoE large models during reinforcement learning (RL) training, enabling its advanced reasoning capabilities to steadily improve even over extended training periods. Ring-flash-2.0 has achieved remarkable breakthroughs across multiple demanding benchmarks, including math competitions, code generation, and logical reasoning—outperforming state-of-the-art dense models with fewer than 40B parameters while rivaling larger-scale open-source MoE models and proprietary high-performance reasoning systems. Despite its focus on complex reasoning tasks, the model also excels in creative writing and other applications. Moreover, thanks to its efficient architectural design, Ring-flash-2.0 delivers robust performance alongside rapid inference speeds, significantly reducing the deployment costs of reasoning models in high-concurrency scenarios.