Netease Youdao has recently unveiled a significant innovation by officially launching the country's first reasoning model capable of providing step-by-step explanations - named "Zi Yue-O1". This model will be made open source to the public, marking another crucial milestone for Netease Youdao in the field of AI education.
According to official statements from Netease Youdao, "Zi Yue-O1" is a lightweight single model with only 14 billion parameters, allowing it to run smoothly on consumer-grade graphics cards. The model employs advanced chain-of-thought technology to offer users detailed problem-solving processes. Its robust logic and reasoning capabilities have achieved higher accuracy in solving problems. Notably, "Zi Yue-O1" has been specifically optimized for Chinese logical reasoning, making it more suitable for domestic users' needs.
Netease Youdao points out that there are limited applicable open-source models available on the market today, most of which have large parameter sizes and cannot operate on ordinary consumer-grade graphics cards. Even if some models use low-bit quantization techniques enabling deployment on a single card, this often compromises the stability of long chain-of-thought operations. To address this issue, "Zi Yue-O1" utilizes a smaller base model, allowing for single-card deployment while maintaining superior mathematical capabilities. Through further optimization, "Zi Yue-O1" can deliver cloud-quality model services on consumer-grade graphics cards.
While reducing its scale, "Zi Yue-O1" does not compromise on reasoning abilities. It adopts unique chain-of-thought technology to create the first domestic model capable of outputting step-by-step problem explanations. Despite having just 14 billion parameters, "Zi Yue-O1" replicates the single-model inference capabilities of OpenAI’s o1, which is undoubtedly an impressive achievement within the industry.
Netease Youdao stated that the launch and open-sourcing of "Zi Yue-O1" aim to provide a powerful yet convenient tool for developers, educators, and learners alike, supporting their exploration and practical application in the realm of AI education.
