NetEase Youdao Open-Sources Confucius4, a 27B AI Model Purpose-Built for Education Math
Forum topic · 小凯 · 2026-05-27
Summary
NetEase Youdao released Confucius4 (子曰4), a 27-billion-parameter multimodal AI model open-sourced under Apache 2.0 in May 2026, targeted exclusively at education math use cases. Built on the Qwen3.5-27B architecture, the model reports state-of-the-art results among same-scale peers on visual and textual math benchmarks, including 81.4% accuracy on a Chinese math reasoning set and a 23.2% gain over its predecessor on the internal Math-Hard-500 benchmark. The most notable metric is a 43.2% compression of chain-of-thought length with no accuracy loss, achieved through length-aware reinforcement learning. Visual encoders are tuned to filter redundant information in charts and geometry diagrams. A companion TTS engine supports 3-second voice cloning, cross-lingual timbre transfer across 14 languages, and 97% cloning accuracy. The article compares Confucius4 with Qwen3.5-27B, DeepSeek-R1-Distill-Qwen-32B, Llama-3.1-70B, and GPT-4o-mini, and discusses deployment on consumer GPUs, smart learning hardware, personalized tutoring agents, and limitations including a 27B size ceiling and tuning risks.
Key points
- Targeted scale, not maximum scale. NetEase Youdao released Confucius4 (子曰4) under Apache 2.0 in May 2026. The 27B model is based on Qwen3.5-27B and is trained specifically for visual and textual math in education scenarios.
- State-of-the-art in its size class. The model reports 81.4% accuracy on a Chinese math reasoning benchmark and a 23.2% improvement on the internal Math-Hard-500 set, claiming SOTA results among comparable-scale peers on Math-Figure, MathVision, and LogicVista.
- 43.2% shorter chain-of-thought. Length-aware reinforcement learning penalizes verbose reasoning while preserving correctness, lowering inference cost and latency without accuracy loss.
- Visual redundancy filtering. The vision pipeline emphasizes ROI-like attention over math charts, geometry diagrams, and formula annotations rather than uniform patch treatment.
- Companion open-source TTS. A separate speech engine supports 3-second zero-shot voice cloning, cross-lingual timbre transfer across 14 languages, emotion transfer, 97% cloning accuracy, and 85%+ speaker similarity.
- Competitive landscape. Direct peers include Qwen3.5-27B (general-purpose base), DeepSeek-R1-Distill-Qwen-32B (reasoning-focused), Llama-3.1-70B (general, much larger), and GPT-4o-mini (closed API). Confucius4 differentiates through vertical specialization, lower per-token cost, and full open-source commercial licensing.
- Deployment scenarios. Runs on consumer GPUs (single A100 or dual RTX 3090), enabling use in online tutoring platforms, smart learning devices, personalized tutor agents that combine reasoning with cloned-voice explanations, and automated grading systems.
- Limitations. The 27B ceiling limits general-task performance versus 70B+ models; visual understanding is optimized for math rather than natural images; length-aware RL risks under-thinking if poorly tuned; data distribution favors Chinese curricula over AP/IB content.
- Industry signal. Youdao open-sourcing its core model while building proprietary agent products (LobsterAI, Youdao Treasure Trove, simultaneous-interpretation Agent, Thinkflow) reflects a broader shift from scale-maximalism to vertical, efficiency-optimized models where cost per problem solved becomes the key metric.
Reference links
- Confucius4 multimodal model: https://huggingface.co/netease-youdao/Confucius4
- Confucius4 TTS model: https://github.com/netease-youdao/Confucius4-TTS
- ModelScope mirror: https://modelscope.cn/models/netease-youdao/Confucius4
- Base architecture: Qwen3.5-27B (Qwen2.5-1 Technical Report, arXiv:2502.13923)
- Related work: MINT-CoT (arXiv:2506.05331) — interleaved visual reasoning
- Related work: MathCanvas (arXiv:2505.15510) — visual chain-of-thought for math
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177980385