Arigatou Ziyou-o1 inference model released! First output step-by-step explanation that can be deployed on consumer-grade graphics cards

In the opening year of 2025, the AI industry set off a large model "reasoning wave", since OpenAI released o1, various reasoning models continue to emerge, the model's high-level reasoning ability ushered in an outbreak of enhancement of the value of its application is also more and more to obtain the industry's widespread attention.

Recently, NetEase has officially launchedThe first reasoning model in China that outputs step-by-step explanations"Ziyi-o1". As a 14B lightweight single model, ZiYin-o1 supports deployment on consumer-grade graphics cards and uses thought chain technology to be able to provide a detailed problem solving process to achieve higher problem solving accuracy with strong logic and reasoning capabilities, as well as provide Chinese logical reasoning, which will help the wide application and innovation of reasoning models in the education field.

Currently, Ziyi-o1 is officially open-sourced to the public and its results are already onlineBeginning Intelligence AI-wisemodel open source community.Welcome to go ahead and experience and use it.

model address

https://wisemodel.cn/models/Netease_Youdao/Confucius-o1-14B

1. Introduction to the model

Looking at the current "reasoning wave", longer thinking links to achieve stronger logic and reasoning ability has become the main technical ideas of reasoning models, under the guidance of which models with different characteristics have emerged.

In this case, the available open-source models are but few, and the parameter scale is large, unable to run on low graphics memory consumer-grade graphics cards, even though low-bit quantization technology is used to enable it to be deployed on a single card, but accordingly, it also brings instability to the operation of a long chain of thought. To address this problem, the Ziyi-o1 open-source model chooses a smaller parameter scale base model that is capable of single-card deployment and has stronger mathematical capabilities. Based on this, the Ziyi-o1 open source model is further lightweight, able to run on consumer-grade graphics cards, and provides model quality comparable to that of cloud deployments. In the scale of "compression" at the same time, Ziyi-o1 using the chain of thought technology, to create the first output step-by-step lecture thinking chain model, with 14B small parameter scale can reproduce the OpenAI o1 single model reasoning ability.

Ziyi-o1 will form a longer chain of thought when solving the problem, making its operation idea closer to the human way of thinking, and providing step-by-step problem solving process and final result by "talking to itself" and self-correcting errors.

As an educational pendant model, this feature of Ziyi-o1 is also more compatible with educational application products, which guide students to realize the enhancement of their independent thinking ability by clearly presenting the organized problem solving process with inspirational explanations.

Not only that, for the application in the field of education, Ziyi-o1 is further optimized in terms of data screening and training instructions on the high accuracy achieved by the long thinking chain. By applying the automated assessment method developed by Yudao, Ziyi-o1 not only assesses the correctness of the final answer, but also covers the entire explanation process to ensure the high quality of the learning data.

In the selection of training instructions, based on the accumulation of data resources in the field of education for many years by Yodao, Ziyi-o1 uses a large number of students' test paper exercises in the field of education as training samples, so as to improve the accuracy of the application of educational scenarios.

At present, Ziyi-o1 has been applied in NetEase Youdao's AI general learning assistant "Youdao Xiaopi", supporting it to realize the question-answering process of "providing analytical ideas and then providing answers", guiding the student users to think actively and call upon their knowledge reserves to solve problems on their own. The answer process guides students to think actively and call on their knowledge reserves to solve problems on their own, thus realizing that they can really learn knowledge thoroughly.

In the lightweight, output step-by-step explanation, Chinese logical reasoning and other diversified advantages, ZiYi -o1 can further empower the domestic AI education application to improve quality and efficiency, with a lower landing threshold to pry higher application value.

Adhering to the "scene first", Youdao Ziyi education big model as a big model of education pendant class, already has a more professional pre-training corpus than the general big model, can be based on the user's needs in the learning scene, to help users to answer questions and solve puzzles.

Along with the continuous expansion of the inference model track, NetEase Youtube's deep cultivation and precipitation within the education vertical will also empower its in-depth exploration of the education pendant model, taking ZiYin-o1 as a starting point, and continuing to unleash the value of the application of inference models in the education field.

2.Online experience

The Wisemodel community supports the creation of online experiences directly through model mirroring, by clicking on the "Deploy Online - Deploy Online Experience" button on the model details page of child-o1-14B.

Then choose and confirm the billing method, here choose the billing method of per volume - manual stop, you can choose the billing method according to your needs.

After submitting the order, the application has entered the state of the startup, under normal circumstances, waiting about 1 minute or so on the normal operation of the online experience.