零一万物（01.AI）是一家由知名计算机科学家和投资人李开复创立的人工智能公司，成立于2023年5月。公司专注于AI 2.0技术及人工智能大模型的研发，致力于推动自然语言处理和大模型技术的发展。
Yi 系列模型
Yi-1.5 系列
Yi-1.5 是对原始 Yi 模型的升级版本，具有更强的性能，特别是在代码生成、数学推理、复杂推理和指令遵循等方面。
- 版本：
- Yi-1.5-34B：34亿参数，适用于高复杂度任务。
- Yi-1.5-9B：9亿参数，适用于中等复杂度任务。
- Yi-1.5-6B：6亿参数，适用于较低复杂度任务。
- 特点：
- 预训练数据：使用5000亿高质量语料进行预训练。
- 微调数据：使用300万多样化的微调样本进行微调。
- 应用场景：语言理解、常识推理、阅读理解、代码生成、数学推理等。
Yi-Large
Yi-Large 是零一万物推出的千亿参数闭源大模型，性能在多个评测中表现优异，部分指标超越了GPT-4。
- 参数规模：千亿参数。
- 性能：在通用能力、代码生成、数学推理、指令遵循等多方面表现出色。
- 应用场景：支持问答、文档解读、PPT制作等多种办公需求。
Yi-XLarge MoE
Yi-XLarge MoE 是零一万物正在训练的更大规模模型，目标是达到甚至超越GPT-5的性能。
- 特点：
- 架构：基于MoE（Mixture of Experts）架构。
- 性能：初步训练结果显示，其各项指标已经超过Yi-Large。
其他模型
零一万物还推出了其他一些专门针对不同应用场景的模型，如Yi-VL（多模态模型）和Yi-Coder（代码生成模型）。
Yi-VL
- 版本：
- Yi-VL-34B：34亿参数。
- Yi-VL-6B：6亿参数。
- 特点：支持多模态输入，能够处理图像并进行相关讨论。
Yi-Coder
- 特点：专注于代码生成，参数规模小于10亿，但在代码生成任务中表现出色。
主要收费模式
1. API 调用收费
零一万物的主要收费模式之一是基于API调用的计费方式。这种模式通常根据用户调用API的次数或处理的token数量进行收费。
- Yi-Large API：目前的收费标准为每百万个token 20元人民币（约合2.7美元）。
- 性价比优势：李开复在媒体专访中强调，Yi-Large模型的付费API性价比优势显著。例如，输入100万个token仅需2.5美元，输出100万个token仅需12美元，相比之下，OpenAI的GPT-4收费更高。
办公场景
1. 会议纪要
零一万物的AI工具可以自动生成会议纪要，帮助用户快速记录和整理会议内容。
- 功能：自动识别和记录会议中的重要信息，生成结构化的会议纪要。
- 优势：提高会议效率，减少人工记录的时间和错误。
2. 周报与写作助手
AI工具可以帮助用户撰写周报和其他工作报告，提供写作建议和自动生成内容。
- 功能：根据用户输入的关键词和要求，生成周报、工作报告等文档。
- 优势：节省时间，提高文档质量。
3. 财报解读
零一万物的AI工具可以快速解读财务报告，提取关键数据和信息。
- 功能：自动分析财务报告中的数据，生成简明扼要的解读和总结。
- 优势：帮助用户快速理解复杂的财务信息，做出更明智的决策。
4. PPT 制作
AI工具可以根据用户输入的主题和要求，自动生成PPT大纲和内容。
- 功能：用户只需输入PPT主题，AI工具即可生成大纲，并在短时间内完成PPT制作。
- 优势：大幅减少PPT制作时间，提高工作效率。
教育场景
1. 智能辅导
AI技术可以用于学生的智能辅导，提供个性化的学习建议和指导。
- 功能：根据学生的学习情况和需求，生成个性化的学习计划和辅导内容。
- 优势：提高学习效率，帮助学生更好地掌握知识。
2. 自动批改作业
AI工具可以自动批改学生的作业，提供详细的反馈和评分。
- 功能：自动识别和批改作业中的错误，生成详细的反馈报告。
- 优势：减轻教师的工作负担，提高批改效率和准确性。
金融场景
1. 风险评估
AI技术可以用于金融风险评估，帮助金融机构识别和管理风险。
- 功能：分析大量金融数据，生成风险评估报告和建议。
- 优势：提高风险管理的准确性和效率。
2. 投资分析
AI工具可以帮助投资者进行投资分析，提供投资建议和预测。
- 功能：分析市场数据和趋势，生成投资建议和预测报告。
- 优势：帮助投资者做出更明智的投资决策。
其他应用场景
1. 客户支持
AI技术可以用于自动化客户支持，提供即时的客户服务和问题解决方案。
- 功能：通过聊天机器人和自动回复系统，提供24/7的客户支持。
- 优势：提高客户满意度，减少人工客服的工作量。
2. 个性化购物体验
AI技术可以用于电子商务平台，提供个性化的购物建议和体验。
- 功能：根据用户的购物历史和偏好，生成个性化的商品推荐。
- 优势：提高用户的购物体验和平台的销售额。
开源版本
Yi-1.5 系列
零一万物的Yi-1.5系列是其开源模型的代表，具有不同参数规模，适用于多种应用场景。
- 版本：
- Yi-1.5-34B：34亿参数，适用于高复杂度任务。
- Yi-1.5-9B：9亿参数，适用于中等复杂度任务。
- Yi-1.5-6B：6亿参数，适用于较低复杂度任务。
- 特点：
- 预训练数据：使用5000亿高质量语料进行预训练。
- 微调数据：使用300万多样化的微调样本进行微调。
- 应用场景：语言理解、常识推理、阅读理解、代码生成、数学推理等。
- 许可证：采用Apache 2.0许可证，允许用户自由使用、修改和分发。
Yi-VL 系列
Yi-VL系列是多模态开源模型，支持图像和文本的联合处理。
- 版本：
- Yi-VL-34B：34亿参数。
- Yi-VL-6B：6亿参数。
- 特点：支持多模态输入，能够处理图像并进行相关讨论。
Yi-Coder 系列
Yi-Coder系列专注于代码生成任务，尽管参数规模较小，但在代码生成任务中表现出色。
- 特点：专注于代码生成，参数规模小于10亿，但在代码生成任务中表现出色。
闭源版本
Yi-Large
Yi-Large是零一万物推出的千亿参数闭源大模型，性能在多个评测中表现优异，部分指标超越了GPT-4。
- 参数规模：千亿参数。
- 性能：在通用能力、代码生成、数学推理、指令遵循等多方面表现出色。
- 应用场景：支持问答、文档解读、PPT制作等多种办公需求。
- 收费模式：基于API调用的计费方式，每百万个token 20元人民币（约合2.7美元）。
Yi-XLarge MoE
Yi-XLarge MoE是零一万物正在训练的更大规模模型，目标是达到甚至超越GPT-5的性能。
- 特点：
- 架构：基于MoE（Mixture of Experts）架构。
- 性能：初步训练结果显示，其各项指标已经超过Yi-Large。
Zero One Technology (01.AI) is an AI company founded by renowned computer scientist and investor Kai-Fu Lee in May 2023. The company focuses on AI 2.0 technologies and the development of large-scale AI models, aiming to advance natural language processing and large-model technologies.
Yi Series Models
- Yi-1.5 Series
Yi-1.5 is an upgraded version of the original Yi model, offering enhanced performance, particularly in code generation, mathematical reasoning, complex reasoning, and instruction-following tasks.
Versions:
- Yi-1.5-34B: 34 billion parameters, suited for high-complexity tasks.
- Yi-1.5-9B: 9 billion parameters, suited for medium-complexity tasks.
- Yi-1.5-6B: 6 billion parameters, suited for lower-complexity tasks.
Features:
- Pre-training data: Trained on 500 billion high-quality tokens.
- Fine-tuning data: Fine-tuned using 3 million diversified samples.
- Applications: Language understanding, commonsense reasoning, reading comprehension, code generation, and mathematical reasoning.
- Yi-Large
Yi-Large is a closed-source model with hundreds of billions of parameters. It has excelled in multiple benchmarks, surpassing GPT-4 in some metrics.
- Parameter size: Hundreds of billions of parameters.
- Performance: Strong performance in general capabilities, code generation, mathematical reasoning, and instruction-following.
- Applications: Supports Q&A, document interpretation, PPT creation, and other office-related needs.
- Yi-XLarge MoE
Yi-XLarge MoE is a larger model currently in training, aiming to achieve or surpass GPT-5-level performance.
Features:
- Architecture: Based on the Mixture of Experts (MoE) architecture.
- Performance: Initial training results show that its metrics already surpass Yi-Large.
Other Models
- Yi-VL (Multimodal Model)
The Yi-VL series is designed to handle multimodal input, processing both images and related discussions.
Versions:
- Yi-VL-34B: 34 billion parameters.
- Yi-VL-6B: 6 billion parameters.
- Yi-Coder
Yi-Coder focuses on code generation tasks, with a parameter size of less than 1 billion, but delivers outstanding performance in code-related tasks.
Main Pricing Model
- API Call Charges
One of the primary billing methods for Zero One Technology is based on API calls. Fees are usually charged according to the number of API calls or the number of tokens processed.
- Yi-Large API: The current rate is ¥20 per million tokens (approximately $2.7 USD).
- Cost Efficiency: In a media interview, Kai-Fu Lee emphasized the significant cost-efficiency advantage of the Yi-Large model’s API. For instance, 1 million input tokens cost only $2.5, and 1 million output tokens cost only $12, which is more cost-effective than OpenAI’s GPT-4.
Office Applications
- Meeting Summaries
Zero One Technology’s AI tools can automatically generate meeting summaries, helping users quickly record and organize meeting content.
Features: Automatically identifies and records key information from meetings, generating structured summaries. Benefits: Improves meeting efficiency and reduces manual note-taking time and errors.
- Weekly Reports & Writing Assistant
AI tools can assist users in writing weekly reports and other work documents by providing writing suggestions and auto-generating content.
Features: Generates weekly reports and work documents based on user-provided keywords and requirements. Benefits: Saves time and improves document quality.
- Financial Report Interpretation
The AI tools can quickly interpret financial reports, extracting key data and information.
Features: Automatically analyzes data from financial reports, generating concise interpretations and summaries. Benefits: Helps users quickly understand complex financial information for better decision-making.
- PPT Creation
AI tools can automatically generate PPT outlines and content based on user input.
Features: Users input the PPT theme, and the AI tool generates an outline and completes the PPT in a short time. Benefits: Significantly reduces PPT creation time and improves work efficiency.
Educational Applications
- Intelligent Tutoring
AI technology can be used to provide personalized learning suggestions and guidance to students.
Features: Generates personalized learning plans and tutoring content based on the student’s progress and needs. Benefits: Increases learning efficiency and helps students better grasp knowledge.
- Automated Homework Grading
AI tools can automatically grade students’ homework, providing detailed feedback and scores.
Features: Automatically identifies errors in homework and generates detailed feedback reports. Benefits: Reduces teachers’ workload and improves grading efficiency and accuracy.
Financial Applications
- Risk Assessment
AI technology can be used for financial risk assessment, helping financial institutions identify and manage risks.
Features: Analyzes large volumes of financial data, generating risk assessment reports and recommendations. Benefits: Improves the accuracy and efficiency of risk management.
- Investment Analysis
AI tools assist investors in conducting investment analysis, offering investment suggestions and forecasts.
Features: Analyzes market data and trends, generating investment recommendations and prediction reports. Benefits: Helps investors make more informed investment decisions.
Other Applications
- Customer Support
AI technology can be used for automated customer support, providing instant solutions to customer queries.
Features: Provides 24/7 customer support through chatbots and automated response systems. Benefits: Improves customer satisfaction and reduces the workload of human support staff.
- Personalized Shopping Experience
AI technology can enhance e-commerce platforms by offering personalized shopping suggestions and experiences.
Features: Generates personalized product recommendations based on users’ shopping history and preferences. Benefits: Improves user shopping experience and increases platform sales.
Open-Source Versions
- Yi-1.5 Series
The Yi-1.5 series is Zero One Technology’s open-source model, with various parameter scales for different application scenarios.
Versions:
- Yi-1.5-34B: 34 billion parameters, suited for high-complexity tasks.
- Yi-1.5-9B: 9 billion parameters, suited for medium-complexity tasks.
- Yi-1.5-6B: 6 billion parameters, suited for lower-complexity tasks.
Features:
- Pre-training data: Trained on 500 billion high-quality tokens.
- Fine-tuning data: Fine-tuned using 3 million diversified samples.
- Applications: Language understanding, commonsense reasoning, reading comprehension, code generation, and mathematical reasoning.
- License: Licensed under Apache 2.0, allowing free use, modification, and distribution.
- Yi-VL Series
The Yi-VL series is an open-source multimodal model that supports joint processing of images and text.
Versions:
- Yi-VL-34B: 34 billion parameters.
- Yi-VL-6B: 6 billion parameters.
Features: Supports multimodal input, processing both images and related discussions.
- Yi-Coder Series
The Yi-Coder series specializes in code generation tasks. Despite its smaller parameter size, it excels in code generation tasks.
Features: Focuses on code generation, with less than 1 billion parameters, but delivers outstanding performance in code-related tasks.
Closed-Source Versions
- Yi-Large
Yi-Large is Zero One Technology’s closed-source model with hundreds of billions of parameters. It excels in multiple benchmarks, surpassing GPT-4 in some metrics.
- Parameter size: Hundreds of billions of parameters.
- Performance: Outstanding in general capabilities, code generation, mathematical reasoning, and instruction-following.
- Applications: Supports Q&A, document interpretation, PPT creation, and other office-related needs.
- Pricing model: API call charges, ¥20 per million tokens (approximately $2.7 USD).
- Yi-XLarge MoE
Yi-XLarge MoE is a larger model in training, with the goal of achieving or surpassing GPT-5-level performance.
Features:
- Architecture: Based on the Mixture of Experts (MoE) architecture.
- Performance: Initial training results show that its metrics have already surpassed Yi-Large.
