The Ant Bai Ling large model team recently announced the official open-source release of its latest high-performance reasoning model — Ring-flash-2.0. This model is a deeply optimized version of Ling-flash-2.0-base, marking a significant advancement in the field of artificial intelligence. Ring-flash-2.0 has a total of 10 billion parameters, but only 610 million are activated during each inference, providing powerful computing capabilities while saving a significant amount of resources through this efficient activation mechanism.
According to the team's introduction, Ring-flash-2.0 performs exceptionally well in multiple challenging benchmark tests, including math competitions, code generation, and logical reasoning. Its performance not only surpasses similar 4 billion parameter models, but also competes with larger-scale open-source sparse models (MoE) and some closed-source high-performance reasoning model APIs, demonstrating its outstanding competitiveness.
To comprehensively enhance the capabilities of Ring-flash-2.0, the Ant Bai Ling team designed an innovative two-phase reinforcement learning (RL) training process. First, through lightweight Long-CoT (long sequence chain-of-thought) SFT (supervised fine-tuning), the Ling-flash-2.0-base model is able to master various thinking methods. Then, RLVR (reinforcement learning with verifiable rewards) training is used to continuously stimulate the model's reasoning potential. Finally, the RLHF (reinforcement learning with human feedback) phase is added to enhance the model's general capabilities.
Notably, the model weights, reinforcement learning training plan, and data recipes of Ring-flash-2.0 will be fully open-sourced, providing valuable resources for developers and researchers. Interested users can access the relevant materials on Hugging Face and ModelScope to start exploring this powerful model.
With the continuous development of AI technology, Ring-flash-2.0 undoubtedly opens up new possibilities for future intelligent applications. We look forward to its widespread application and further breakthroughs in various fields!
Model Address:
https://huggingface.co/inclusionAI/Ring-flash-2.0
https://modelscope.cn/models/inclusionAI/Ring-flash-2.0
