What is Ring-lite
Ring-lite is a lightweight reasoning model based on the MoE architecture launched by the Ant Technology AntTech team. The model uses Ling-lite-1.5 as its foundation and is based on the original C3PO reinforcement learning training method, achieving SOTA results on multiple reasoning benchmarks with only 2.75B activated parameters. Ring-lite achieves efficient reasoning through technological innovations such as stable reinforcement learning training, optimizing the training proportion of Long-CoT SFT and RL, and solving the challenges of joint training across multiple domains. Ring-lite has open-sourced its technology stack, including model weights, training code, datasets, etc., promoting the development of lightweight MoE reasoning models.
Main Features of Ring-lite
- Efficient Reasoning: Ring-lite can achieve efficient reasoning in multiple complex reasoning tasks, such as mathematical reasoning, programming competitions, and scientific reasoning.
- Lightweight Design: Ring-lite has a total parameter count of 16.8B, with only 2.75B activated parameters. While maintaining high performance, it has low computational resource requirements and is suitable for use in resource-constrained environments.
- Multi-Domain Reasoning: Ring-lite can handle reasoning tasks in multiple domains, including mathematics, programming, and science. Through joint training and phased training methods, it achieves synergistic gains across different domains and improves the model's generalization ability.
- Stable Training: Ring-lite is based on the C3PO reinforcement learning training method, solving the instability issues in traditional reinforcement learning training and improving the stability and efficiency of the training process.
Technical Principles of Ring-lite
- MoE Architecture: Ring-lite is based on the Mixture-of-Experts (MoE) architecture, processing input data based on a combination of multiple expert networks. Each expert network is responsible for handling specific subtasks or data features, improving the model's reasoning ability and efficiency.
- C3PO Reinforcement Learning Training Method: C3PO (Constrained Contextual Computation Policy Optimization) is an innovative reinforcement learning training method. By fixing the total number of training tokens per training step, it eliminates the optimization instability and throughput fluctuations caused by variations in generated response length. C3PO selects an appropriate starting model based on an entropy loss strategy, further stabilizing the training process.
- Combination of Long-CoT SFT and RL: Ring-lite is based on a two-stage training method that combines Long-CoT (long chain-of-thought) supervised fine-tuning (SFT) and reinforcement learning (RL). Through Long-CoT SFT, the model can learn complex reasoning patterns; through RL, the model can further optimize performance on specific tasks. Ring-lite determines the optimal training proportion of SFT and RL based on experiments, achieving a balance between token efficiency and performance.
- Joint Training with Multi-Domain Data: During training, Ring-lite uses data from multiple domains such as mathematics, programming, and science. Based on a phased training method, it trains on mathematical tasks first and then conducts joint training on code and science tasks, effectively solving the domain conflict problem in joint training with multi-domain data.
Project Address of Ring-lite
- GitHub Repository：https://github.com/inclusionAI/Ring
- HuggingFace Model Library：https://huggingface.co/inclusionAI/Ring-lite
- arXiv Technical Paper：https://arxiv.org/pdf/2506.14731
Application Scenarios of Ring-lite
- Education: Assists students in solving complex mathematical and scientific problems, providing detailed solution steps and reasoning processes to help students better understand and master knowledge.
- Scientific Research: Assist researchers in verifying and exploring complex mathematical and scientific problems, provide detailed reasoning steps and solutions, and support theoretical research and experimental design.
- Industry and Commerce: Process and analyze complex data, provide reasoning processes and solutions, for forecasting and decision support in fields such as finance, healthcare, and marketing.
- Intelligent Assistants: Integrate into intelligent assistants to provide more intelligent reasoning and answering capabilities, handle complex queries and tasks, and improve user experience.
- Healthcare: Assist doctors and researchers in analyzing and processing complex medical data, provide detailed reasoning processes and solutions, and support disease diagnosis and treatment plan development.