Youdao (NYSE:DAO) Launches Lightweight Inference Model "Confucius-o1," Achieving Stronger Inference with a Smaller Scale

In 2025, the AI industry has witnessed a surge in the development of large-scale inference models, following OpenAI's release of o1. Various inference models have been emerging, with their high-level reasoning capabilities significantly enhanced and their application value increasingly recognized by the industry.

On January 22, NetEase Youdao officially launched China's first step-by-step exposition inference model, "Confucius-o1." As a 14B lightweight single model, Confucius-o1 supports deployment on consumer-grade GPUs and utilizes chain-of-thought technology to provide detailed problem-solving processes. It achieves higher accuracy in problem-solving with strong logic and reasoning capabilities and offers reasoning in Chinese. Notably, Confucius-o1 has officially been open-sourced to encourage broader use and innovation of inference models in the education sector.

With the rise of the current "reasoning trend," the primary technical approach has been to achieve stronger logic and reasoning capabilities by longer chains of thought, leading to the emergence of various models with distinct characteristics. However, there are few open-source models available, and most have large parameter scales that cannot run on low-memory consumer GPUs. Even with low-bit quantization technology, which allows deployment on a single card, it causes instability to the operation of long chains of thought.

To address this issue, Confucius-o1, as an open-source model, uses a smaller parameter scale base model that can be deployed on a single card and has stronger mathematical capabilities. This allows Confucius-o1 to further achieve lightweighting and run on consumer-grade GPUs while providing model quality comparable to cloud deployment.

Image: https://www.globalnewslines.com/uploads/2025/01/8ec4e8d62b105466930508e847f184dd.jpg

By reducing its scale, Confucius-o1 has adopted chain-of-thought technology to create China's first step-by-step problem-solving chain-of-thought model. With a 14B small parameter scale, it matches the single-model inference capabilities of OpenAI's o1. It forms longer chains of thought during problem-solving, making it process more similar to human thinking. Through "self-talk" and self-correction, it provides step-by-step problem-solving processes and final results. As an education-focused model, this feature of Confucius-o1 is well-suited for educational applications, helping students develop independent thinking skills through clear, logical problem-solving.

Image: https://www.globalnewslines.com/uploads/2025/01/aa08602b4015800e9dcc8bbc87368d9d.jpg

Moreover, targeting educational applications, Confucius-o1 has further optimized its high accuracy achieved through long chains of thought, from data filtering and training instructions. Using Youdao's self-developed automated evaluation method, Confucius-o1 not only assesses the correctness of the final answers but also covers the entire explanation process to ensure high-quality learning data.

In terms of training instruction selection, Confucius-o1 leverages Youdao's extensive data resources accumulated over the years in the education sector, using a large number of student exam questions and practice problems as training samples to improve accuracy in educational scenarios.

Currently, Confucius-o1 has been implemented in NetEase Youdao's AI All-in-One Learning Assistant, "Mr. P AI Tutor," supporting its "first provide the problem-solving, then provide the answer" process. This guides student users to think actively and use their knowledge reserves to solve problems independently, truly mastering the knowledge. With its lightweight, step-by-step explanation, and Chinese logic reasoning advantages, Confucius-o1 can further empower domestic AI educational applications to improve quality and efficiency, leveraging lower deployment thresholds to unlock higher application value.

Adhering to the principle of "scene first," confucius large language model, as a vertical education model, already has more professional pre-training corpora than general large models. It can help users solve problems based on their learning needs. As the inference model track continues to expand, NetEase Youdao's in-depth exploration in the education vertical field will continue to empower its exploration of education vertical models, starting with Confucius-o1 to continuously unlock more value in education through inference models.

Welcome to visit the Demo address for experience: https://confucius-o1-demo.youdao.com/

Download Links for the Model:

https://huggingface.co/netease-youdao/Confucius-o1-14B

https://modelscope.cn/models/netease-youdao/Confucius-o1-14B
Media Contact
Company Name: Youdao.inc
Contact Person: Licaifeng
Email: Send Email [http://www.universalpressrelease.com/?pr=youdao-nysedao-launches-lightweight-inference-model-confuciuso1-achieving-stronger-inference-with-a-smaller-scale]
Country: China
Website: https://www.youdao.com



This release was published on openPR.