Ring-lite - A lightweight inference model open-sourced by Ant Financial.

What is Ring-lite?

Ring-lite is a lightweight inference model based on the MoE architecture, developed by AntTech. Based on Ling-lite-1.5, the model utilizes the proprietary C3PO reinforcement learning training method, achieving state-of-the-art (SOTA) performance on multiple inference benchmarks with only 2.75B activation parameters. Ring-lite achieves efficient inference through technological innovations such as stable reinforcement learning training, optimized training weights for Long-CoT SFT and RL, and solutions to the challenges of joint training across multiple domains. Ring-lite has open-sourced its technology stack, including model weights, training code, and datasets, contributing to the development of lightweight MoE inference models.

Ring-lite's main functions

• Efficient ReasoningRing-lite enables efficient reasoning in a variety of complex reasoning tasks, such as mathematical reasoning, programming competitions, and scientific reasoning.
• Lightweight designRing-lite has a total parameter count of 16.8B and an activation parameter count of only 2.75B. While maintaining high performance, it has low computational resource requirements and is suitable for use in resource-constrained environments.
• Multi-domain reasoningRing-lite can handle reasoning tasks across multiple domains, including mathematics, programming, and science. It achieves synergistic gains across different domains through joint training and phased training methods, improving the model's generalization ability.
• Stable trainingRing-lite is based on the C3PO reinforcement learning training method, which solves the instability problem in traditional reinforcement learning training and improves the stability and efficiency of the training process.

Ring-lite's technical principles

• MoE architectureRing-lite is based on the Mixture-of-Experts (MoE) architecture, which processes input data through a combination of multiple expert networks. Each expert network is responsible for processing specific sub-tasks or data features, improving the model's inference ability and efficiency.
• C3PO reinforcement learning training methodC3PO (Constrained Contextual Computation Policy Optimization) is an innovative reinforcement learning training method. By fixing the total number of training tokens at each training step, it eliminates optimization instability and throughput fluctuations caused by variations in the generated response length. C3PO uses an entropy loss strategy to select a suitable starting model, further stabilizing the training process.
• Combination of Long-CoT SFT and RLRing-lite employs a two-stage training approach combining Long-CoT (Long Inference Chain) supervised fine-tuning (SFT) and reinforcement learning (RL). Using Long-CoT SFT, the model learns complex inference patterns; based on RL, the model can further optimize performance on specific tasks. Ring-lite experimentally determines the optimal ratio of SFT to RL training, achieving a balance between token efficiency and performance.
• Multi-domain data joint trainingRing-lite utilizes data from multiple domains, including mathematics, programming, and science, during its training process. Based on a phased training approach, it trains on mathematical tasks and then performs joint training on coding and scientific tasks, effectively resolving domain conflict issues in joint training of multi-domain data.

Ring-lite project address

• GitHub repositoryhttps://github.com/inclusionAI/Ring
• HuggingFace model libraryhttps://huggingface.co/inclusionAI/Ring-lite
• arXiv technical paper: https://arxiv.org/pdf/2506.14731

Applications of Ring-lite

• EducationIt 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 fieldIt assists researchers in verifying and exploring complex mathematical and scientific problems, providing detailed reasoning steps and solutions to support theoretical research and experimental design.
• Industrial and commercial sectorsIt processes and analyzes complex data, providing reasoning processes and solutions for forecasting and decision support in fields such as finance, healthcare, and marketing.
• Smart AssistantIt can be integrated into the intelligent assistant to provide more intelligent reasoning and problem-solving capabilities, handle complex queries and tasks, and improve the user experience.
• medical fieldIt assists doctors and researchers in analyzing and processing complex medical data, providing detailed reasoning processes and solutions to support disease diagnosis and treatment planning.