Confucius4 - NetEase Youdao's Open-Source Multimodal Reasoning Model

What is Confucius4

Confucius4 is an open-source multimodal large model launched by the NetEase Youdao AI team based on the Qwen3.5-27B architecture, specifically designed for advanced mathematical reasoning scenarios. Through an iterative SFT+RL training paradigm, pure-text reasoning data augmentation, and refined CoT optimization, the model achieves SOTA performance among models of the same scale, while reducing reasoning chain length by 43.2%. The model is specifically optimized for Chinese users, producing output that better conforms to Chinese language conventions and cultural background, and supports free commercial use.

Confucius4's Main Features

• Advanced Multimodal Mathematical Reasoning: Supports mixed image-text input, excels at solving complex mathematical problems in geometry, algebra, logic, etc., and performs excellently on benchmarks such as Math-Hard-500, MathVision, and logicVista.
• Iterative SFT+RL Training Optimization: Uses image gain filtering to build a cost-effective training set, and continuously improves performance in text and multimodal scenarios through iterative supervised fine-tuning and reinforcement learning.
• Pure-Text Reasoning Enhancement: Injects pure-text reasoning data during the SFT stage to strengthen the model's reasoning foundation, improving Math-Hard-500 performance by 23.2%.
• Compact Chain-of-Thought Generation: Through refined CoT reconstruction and a length-aware RL mechanism, eliminates redundant reasoning steps and achieves a balance between accuracy and efficiency.
• Chinese-Specific Optimization: Specifically trained on Chinese data, producing content that better fits Chinese users' expression habits and cultural context.
• Open Source and Commercially Usable: Released under the Apache 2.0 license, supporting free modification, distribution, and commercial applications.

Confucius4's Technical Principles

• Image Gain Filtering: Automatically identifies and filters low-value visual redundant information, builds a cost-effective multimodal training dataset, and reduces training costs.
• Iterative SFT+RL Paradigm: Alternates between supervised fine-tuning and reinforcement learning, forming a "training-evaluation-optimization" closed loop, continuously raising the model's reasoning ceiling on text and multimodal tasks.
• Hybrid Training Strategy: Adopts a hybrid training paradigm of "text reasoning + multimodal problem solving," enabling pure-text reasoning capabilities to transfer to multimodal scenarios and achieving synergistic capability gains.
• Refined CoT Reconstruction: Manually reconstructs the chain of thought during the SFT stage, removes redundant steps, retains complete logic, and generates concise and high-quality reasoning chains.
• Length-Aware Advantage: Introduces an exploration-exploitation trade-off during the RL stage, constrains reasoning length for non-hard problems, and effectively eliminates the "overthinking" phenomenon.

How to Use Confucius4

• Environment Preparation: Confirm that the runtime environment meets the Qwen3.5 model requirements, and installtransformersand other necessary dependency libraries.
• Load Model: UseAutoModelForCausalLMandAutoProcessorto load from HuggingFace or ModelScopenetease-youdao/Confucius4pretrained model.
• Encode Image: For multimodal input, usebase64to encode the target image into data URI format.
• Construct Messages: Assemble conversation messages according to the system prompt template, with the system role fixed asYou are a helpful assistant.。
• Apply Template: Callprocessor.apply_chat_templateto process messages and generate input text usable by the model.
• Model Inference: SetTemperature=0.6、TopP=0.95、TopK=20, callmodel.generateto generate results.
• Parse Output: When calling via the vLLM API, obtain the final answer frommessage.content, and obtain the reasoning process frommessage.reasoning.

Core Advantages of Confucius4

• SOTA Performance at the Same Scale: Leads models of the same scale on multiple visual math benchmarks such as Math-Hard-500 (0.814), Math-Figure (0.907), and MathVision (0.724).
• Significantly Improved Reasoning Efficiency: CoT length is reduced by 43.2%, greatly lowering the number of output tokens while maintaining accuracy, reducing reasoning time and computational cost.
• Balanced Chinese and English Scenarios: The model has strong English mathematical reasoning capabilities and, through targeted optimization with Chinese data, produces output that better matches the reading habits of local users.
• Seamless Compatibility with the Qwen Ecosystem: Environment requirements are fully consistent with Qwen3.5, and it can be loaded directly with Transformers or vLLM, with zero additional adaptation cost.

Project Address of Confucius4

• HuggingFace Model Repository：https://huggingface.co/netease-youdao/Confucius4

Comparison of Confucius4 with Similar Competing Products

Application Scenarios of Confucius4

• K12 and Higher Education Mathematics Tutoring: Supports mixed image-text input, solves complex mathematics problems in geometry, algebra, probability, and more, and provides interpretable step-by-step reasoning, making it suitable as an intelligent teaching assistant.
• Mathematics Competitions and Olympiad Training: Achieves an accuracy of 81.4% on competition-level hard problems (Math-Hard-500), making it suitable for training on high-difficulty problem types and analyzing problem-solving strategies.
• Intelligent Question Bank Analysis and Homework Grading: Automatically recognizes image-based questions in exam papers and workbooks, generates complete reasoning chains and final answers, and assists teachers in efficient grading.
• Integration with Educational Hardware and Online Learning Platforms: Compatible with the Qwen3.5 ecosystem, it can be quickly integrated into learning devices and education apps, providing low-latency real-time problem-solving services through the vLLM API.
• Scientific Research and Academic Assistance: Parses mathematical charts, formula derivations, and logical proofs in papers, helping researchers quickly understand visual mathematical content in complex academic materials.