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

1. What Is Confucius4

Confucius4 is an open-source multimodal large model from NetEase Youdao's AI team, built on the Qwen3.5-27B architecture and designed for advanced mathematical reasoning. Through innovative iterative SFT+RL training paradigm, pure-text reasoning data augmentation, and refined chain-of-thought optimization, it achieves SOTA performance among models of comparable scale while reducing reasoning chain length by 43.2%. The model is optimized for Chinese users with output better aligned to Chinese language habits and cultural context, and supports free commercial use.

Image source: Official article

Technical positioning and domain: Confucius4 belongs to multimodal large model domain, focused on visual mathematical reasoning. Core positioning is providing high-precision, high-efficiency mathematical problem-solving for education and research, with significant performance advantages in complex tasks requiring image-text combination such as geometry, algebra, and logical reasoning. Unlike general multimodal models, Confucius4 is deeply optimized for mathematical reasoning, forming differentiated technical barriers.

Development background: NetEase Youdao, a leading Chinese EdTech company, has long focused on intelligent education with rich experience in mathematical reasoning, OCR, and educational AI. Confucius4's motivation stems from urgent demand for high-precision mathematical reasoning models in education—general models often produce logical errors or redundant reasoning on math problems, failing reliability requirements for teaching. Secondary development on Qwen3.5 architecture enabled focused breakthroughs in mathematical reasoning capability.

Core value: Confucius4 solves the "precision-efficiency" trade-off in multimodal mathematical reasoning. Traditional approaches either pursue high precision with lengthy reasoning chains and high compute cost, or sacrifice precision for efficiency. Through refined CoT reconstruction and length-aware RL mechanisms, Confucius4 reduces reasoning chain length by 43.2% while maintaining accuracy, balancing accuracy and reasoning efficiency. Chinese-targeted optimization performs better in Chinese education scenarios, filling a gap in open-source models for this domain.

Technical characteristics: Core advantages include iterative SFT+RL training paradigm with "train-evaluate-optimize" closed loop continuously improving reasoning; image gain filtering automatically identifying and filtering low-value visual redundancy to reduce training cost; length-aware advantage mechanism constraining non-difficult problem reasoning length in RL stage, effectively eliminating "overthinking"; hybrid training strategy migrating pure-text reasoning to multimodal scenarios for synergistic capability gains.

2. Key Features

• Advanced multimodal mathematical reasoning: Supports image-text mixed input, excelling at geometry, algebra, logic, and other complex math problems. Strong performance on Math-Hard-500, MathVision, logicVista, and other benchmarks—especially geometry tasks accurately identifying angle and length relationships in figures and generating complete proof processes. Suited for educational tutoring and competition training requiring visual understanding.

• Iterative SFT+RL training optimization: Builds cost-effective training sets via image gain filtering, continuously improving text and multimodal performance through iterative supervised fine-tuning and reinforcement learning. Forms "train-evaluate-optimize" closed loop ensuring quantifiable improvement on key metrics each iteration, avoiding single-training convergence issues to suboptimal solutions.

• Pure-text reasoning enhancement: Injects pure-text reasoning data during SFT stage to strengthen reasoning foundation. Experiments show 23.2% Math-Hard-500 performance improvement. Migrating pure-text mathematical reasoning to multimodal scenarios yields stronger logical coherence and reasoning depth in image-text mixed tasks.

• Compact chain-of-thought generation: Through refined CoT reconstruction and length-aware RL, eliminates redundant reasoning steps, balancing accuracy and efficiency. Compared to baseline, reasoning chain length reduced 43.2% while maintaining or improving accuracy. Users get faster answers with clearer, more concise reasoning—suited for real-time applications requiring response speed.

• Chinese-targeted optimization: Specialized training on Chinese data with output better aligned to Chinese user expression habits and cultural context. Uses teaching-standard language in Chinese math problem solving, avoiding awkward translation tone or non-Chinese expression patterns, improving user experience.

• Open source, commercially usable: Released under Apache 2.0, supporting free modification, distribution, and commercial application. Developers can build secondary development on Confucius4 integrated into proprietary EdTech products without licensing fees, lowering AI education application barriers especially for small and medium EdTech companies.

3. How to Use

1. Environment preparation: Confirm environment meets Qwen3.5 requirements. Recommend NVIDIA A100 or higher GPU with at least 40GB VRAM. OS: Linux (Ubuntu 20.04+) or Windows Server 2019+. Install Python 3.10+ and dependencies including transformers>=4.45.0, torch>=2.1.0, accelerate, flash-attn, etc.

2. Load model: Use AutoModelForCausalLM and AutoProcessor to load from HuggingFace or ModelScope: model = AutoModelForCausalLM.from_pretrained("netease-youdao/Confucius4", torch_dtype=torch.bfloat16, device_map="auto"). Enable Flash Attention for faster inference: attn_implementation="flash_attention_2".

3. Encode images: For multimodal input, use base64 to encode images as data URI: with open(image_path, "rb") as f: image_base64 = base64.b64encode(f.read()).decode("utf-8"). Keep resolution moderate, recommended max 1024x1024 to avoid VRAM overflow.

4. Construct messages: Assemble conversation messages per system prompt template. System role fixed as You are a helpful assistant. User message contains image and text: {"role": "user", "content": [{"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_base64}"}}, {"type": "text", "text": "Please solve this math problem"}]}.

5. Apply template: Call processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) to generate model input. Set add_generation_prompt=True for generation prompt.

6. Model inference: Set parameters: Temperature=0.6, TopP=0.95, TopK=20. Call model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.6, top_p=0.95, top_k=20). For complex problems, set max_new_tokens to 4096+ for complete reasoning chains.

7. Parse output: Via vLLM API, get final answer from message.content, reasoning from message.reasoning. With Transformers, decode generated_ids and extract model output portion. Filter input prompt, keep generated content only.

8. Notes: First load requires downloading ~54GB parameters—high-speed network recommended. If VRAM insufficient, try device_map="auto" or CPU offloading. For batch inference, recommend vLLM deployment for throughput.

4. Pros and Cons

5. Comparison with Similar Tools

Selection recommendations:

For EdTech companies or research institutions focused on mathematical reasoning, Confucius4 is optimal. Leading benchmark performance and reasoning efficiency significantly improve product competitiveness. Especially for intelligent tutoring handling geometry, algebra, and image-text mixed problems—compact CoT reduces user wait time and improves interaction. Apache 2.0 allows commercial use with lower legal risk.

If general multimodal capability needed (image-text QA, visual description, document understanding simultaneously), consider Qwen2.5-VL-72B or InternVL2-26B—more comprehensive on general tasks but weaker math specialization. For pure-text math (step analysis, proofs), MathPile and similar lightweight models may be more cost-effective but lack visual understanding.

6. Editor's Review

Confucius4 shows clear technical thinking: not blindly pursuing general capability but deep optimization in mathematical reasoning vertical. Iterative SFT+RL, image gain filtering, length-aware advantage mechanisms form a complete technical closed loop. The 43.2% reasoning chain reduction while maintaining accuracy has real application value. Technically, Confucius4 proves medium-scale models (27B) with specialized optimization can match or exceed larger models—important reference for resource-constrained teams.

Practically, Apache 2.0 and Qwen ecosystem compatibility make Confucius4 highly attractive for EdTech open source. 81.4% accuracy on Math-Hard-500 meets most K12 and higher education tutoring needs. Efficiency gains mean lower operating costs and better UX. High hardware requirements (A100 recommended) remain main adoption barrier—awaiting quantized or smaller variants.

Clear target audience: EdTech AI engineers, math researchers, developers needing high-precision mathematical reasoning. Individual learners and light users should wait for community quantized versions or cloud API.

Strong future potential as multimodal AI penetrates education. Recommended continued optimization: smaller variants (7B-14B) lowering hardware barrier; expansion to advanced math and statistics; formula editor support for better interaction.

7. Application Scenarios

• K12 and higher education math tutoring: Image-text input for geometry, algebra, probability, and complex problems. Students photograph and upload questions; model generates explainable step-by-step reasoning. Suited as intelligent teaching assistant for personalized homework or instant Q&A. >90% accuracy on gaokao math in testing.

• Math competition and Olympiad training: 81.4% accuracy on competition-level Math-Hard-500 for high-difficulty training and strategy analysis. Coaches generate reasoning chains for different solution approaches. Rigorous IMO-level reasoning as competition training aid.

• Intelligent question bank parsing and homework grading: Auto-recognize image questions from exams and workbooks, generate complete reasoning chains and final answers. Teachers batch-upload homework photos for auto-grading and error analysis. Compact CoT generates concise feedback, improving efficiency.

• EdTech hardware and online platform integration: Qwen3.5 ecosystem compatible for quick integration into learning devices and education apps. vLLM API provides low-latency real-time solving with concurrent requests. Smart learning devices can integrate Confucius4 as built-in math assistant via voice or photo input.

• Research academic assistance: Parse mathematical charts, formula derivations, and logical proofs in papers. Researchers upload complex formula screenshots for automatic understanding and explanation. Useful for math, physics academic exchange as quick comprehension aid.

8. FAQ

Q: What's the difference between Confucius4 and Qwen3.5-27B?
A: Confucius4 is secondary development on Qwen3.5-27B with iterative SFT+RL, image gain filtering, length-aware advantage mechanisms significantly exceeding original on math reasoning. Math-Hard-500: Confucius4 0.814 vs Qwen3.5-27B 0.582 (+39.9%). Reasoning chain length reduced 43.2%.

Q: What input formats does Confucius4 support?
A: Text and image-text mixed input. Images need base64 data URI encoding; JPEG, PNG supported. Pure text math can use LaTeX formulas or natural language. Recommended max image resolution 1024x1024.

Q: Minimum hardware requirements?
A: Recommend NVIDIA A100 (40GB VRAM) or higher. Minimum 24GB VRAM GPU (RTX 4090) with significantly slower inference. Insufficient VRAM: CPU offloading or quantized versions (community developing). Linux recommended for best performance.

Q: How is Confucius4's Chinese capability vs other models?
A: Specialized Chinese data optimization outperforms same-scale general models on Chinese math. Uses teaching-standard language, avoiding translation tone. Comparable to dedicated Chinese math models on Chinese benchmarks but with unique multimodal input advantage.

Q: How to integrate Confucius4 into my application?
A: Qwen3.5 ecosystem compatible via Transformers or vLLM. Recommend vLLM for production with high concurrency. Steps: install dependencies, load model, implement image encoding, construct message format, call inference API. HuggingFace page has detailed API docs and examples.

9. Project Links

• HuggingFace model hub: https://huggingface.co/netease-youdao/Confucius4
• NetEase Youdao AI team: https://ai.youdao.com/