The race to build bigger general chatbots has a quiet loser: the classroom.
A model that only does schoolwork, it turns out, can out-think giants many times its size on the one job students actually care about.
NetEase's education arm just open-sourced that model — and the math scores are the whole argument.
A model that only does homework
子曰 (Ziyue) is the education-domain large model (大模型) from NetEase Youdao (网易有道), the learning arm of game giant NetEase.
It was launched in July 2023 as, by the company's account, China's first vertical large model built for education. Rather than chase general chat, it was trained around the messy, specific tasks of learning:
- step-by-step problem solving, not just final answers;
- language translation tuned for teaching, not travel phrases;
- tight integration with hardware like dictionary pens and learning tablets.
The point of a vertical model is focus. A generalist spreads its capacity across everything; a classroom model spends its on the curriculum.
The open-weight move
Youdao did not keep 子曰 behind a paywall. Through 2025 it open-sourced the pieces that matter most for research and small deployments.
- 子曰-o1 (Ziyue-o1), released January 2025, is a 14B-parameter reasoning model that replays
OpenAI o1-style chain-of-thought and was described as the first domestic model with explicit step-by-step explanation ability. It is small enough to run on a consumer GPU. - Confucius3-Math (子曰3数学模型), released June 2025, is an open math-education model that runs on a single consumer graphics card. On CK12-math, GAOKAO-Bench(Math), MathBench(K12), and MATH500 it outperformed several larger general-purpose models, per Youdao's release notes.
- 子曰 4.0, announced May 22, 2026, added multimodal input (text, image, audio) and open-sourced its core multimodal and text-to-speech (
TTS) models globally.
NetEase Youdao (网易有道, NYSE: DAO) is a listed company, so the model is backed by real revenue: 2025 net revenue came in at about RMB 59.1亿元 (roughly US$845 million / HK$6.4 billion), up about 5% year on year.
Why small beats big here
A 14B reasoning model losing to a 70B generalist on trivia is expected. Losing on math pedagogy is not, and that is where 子曰-o1 and Confucius3-Math land their hits.
- Education wants explanations, not just answers — a capability that rewards careful reasoning over raw knowledge volume.
- A model that runs on one GPU can sit inside a dictionary pen (词典笔) or a school server without calling a cloud API per question.
- Open weights let schools and EdTech teams audit the model for safety and bias, which closed APIs make harder.
The translation side is no afterthought either: Youdao says 子曰 handles real-time translation across 38 smaller languages, useful for cross-border study material without a massive model.
From a model to a product floor
The model is the engine; the products are the proof. Based on 子曰, Youdao ships:
- the AI answer pen (AI答疑笔) that walks a student through a problem visually, step by step;
- an AI-upgraded dictionary with simultaneous interpretation and document translation;
- classroom tools already deployed across districts in cities like Hangzhou and Beijing.
On standards, 子曰 3.0 passed the China Academy of Information and Communications Technology (中国信通院, CAICT) trusted-AI education-model assessment at its highest 5-level rating in 2025, after a 4+ rating in 2024. For a school buyer, that rating matters more than a leaderboard.
The six-scene classroom
NetEase Youdao frames 子曰 around six education scenarios, which is exactly why a narrow model fits better than a generalist here:
- AI + student learning, where step-by-step help matters more than trivia recall.
- AI + course building for teachers, turning a syllabus into ready materials.
- AI + teacher research, summarizing and structuring academic papers.
- AI + international exchange, leaning on the 38-language translation side.
- AI + campus services, handling routine administrative questions.
- AI + talent development, supporting structured practice and assessment.
The CAICT 5-level rating, the highest tier on China's trusted-AI education-model assessment, signals to school buyers that the model passed an independent education-specific audit — the kind of check a general chatbot leaderboard does not even attempt.
What readers can do now
- Pull Confucius3-Math and test it on your syllabus: for K12 and competition math, a 14B-class open model may beat a much larger closed one on worked-problem quality.
- Use 子曰-o1 for tutoring flows, not chat: its value is the visible reasoning chain, ideal for explanation features.
- If you build EdTech, start from the open weights and add your own safety and curriculum guardrails rather than prompting a general API.
Honest limitations
- The "beats larger general models" claim is sourced from Youdao's own release notes on specific math benchmarks, not an independent head-to-head we ran.
- A 14B education model is narrow by design; it will underperform generalists on broad knowledge and open-domain tasks.
- Running on a consumer GPU is possible but real classroom latency depends on hardware and batch size.
- 子曰 4.0's global open-source of multimodal and TTS weights is recent (May 2026); community adoption and third-party audits are still early.