成立于2021年12月，迅速在AI领域崭露头角。公司自研了文本到视觉、文本到语音以及文本到文本三个基础模型，成为国内首家同时拥有这三种模态大模型能力的创业公司。
ABAB系列模型
ABAB 5.5
- 特点：ABAB 5.5是MiniMax开发的一个大语言模型，具有8,192个token的输入输出限制，适用于字符生成等任务。
- 性能：在写作、聊天、问答等场景中表现良好，达到了GPT-3.5的水平，但在处理更复杂和对输出有精细要求的场景时，存在一定的局限性。
ABAB 6
- 特点：ABAB 6是MiniMax的第二版MoE（Mixture of Experts）大模型，参数量比ABAB 5.5大一个量级，能够更好地处理复杂任务。
- 技术：采用MoE架构，模型参数被划分为多组“专家”，每次推理时只有一部分专家参与计算，从而提升计算效率和模型性能。
- 性能：在多个复杂任务的测试中表现优异，显著提升了模型的推理速度和训练效率。
ABAB 6.5
- 特点：ABAB 6.5是MiniMax推出的万亿参数大语言模型，支持245k上下文窗口，能够满足复杂生产力以及多语言对话场景的需求。
- 性能：在知识、推理、数学、编程、指令遵循等各项测试中接近行业最领先的大模型水平。
按Token计费
- 文本到文本（T2T）：适用于互动娱乐、文本创作等场景。计费单位为每1000个token，约对应750个字符文本（包括标点符号等）。具体价格为每1000个token 0.015元人民币。
- 文本到语音（T2A）：适用于语音播报、语音交互等场景。收费模式以每10000个字符（输入）为基础计价单位，具体字符计算逻辑为1个汉字算2个字符，英文字母、标点符号等算1个字符。
按存储量计费
- 向量存储：检索的向量存储费用按每天的加权平均计算，每GB/天的费用为1.5元人民币。
按会话计费
- Code Interpreter：计价单位为每session 0.2元人民币。一个session定义为1个assistant在1个thread中1个小时内调用Code Interpreter的对话。如果超过1个小时，则算作2个session。
按功能计费
- Chat Completion Pro：在Chat Completion的基础上提供更多功能，包括多人对话、对话示例、插件、函数调用等。每1000个token的价格为0.015元人民币，web search功能每次0.03元人民币。
免费试用
- 免费试用余额：在内测期间，所有账号会统一赠送固定数额的免费试用余额，用户可以免费体验接口调用。正式使用时可通过线上充值，根据实时调用消耗的token数进行计费。
定制服务
- 模型微调：提供定制模型微调服务，价格根据每1000个token的数量、输入文件中的token数量以及训练的epoch数计算。
文本处理
MiniMax的大模型在文本处理方面表现出色，适用于以下场景：
- 逻辑推理：帮助用户进行复杂的逻辑推理任务，提供准确的推理结果。
- 文本续写：自动生成高质量的文本续写内容，适用于小说、文章等创作。
- 文案生成：为广告、营销等场景生成创意文案。
- 内容摘要：从长文本中提取关键信息，生成简洁明了的摘要。
- 代码生成：根据自然语言描述生成相应的代码，提升开发效率。
语音合成
MiniMax的语音模型能够生成自然、高保真、个性化的语音，适用于以下场景：
- 新闻播报：生成专业的新闻播报语音，提升信息传达效果。
- 语音助手：为智能硬件提供语音交互功能，提升用户体验。
- 配音和音频交互：为视频、游戏等提供高质量的配音服务。
智能硬件
MiniMax的大模型在智能硬件领域也有广泛应用：
- 手机语音助手：为手机厂商提供智能语音助手功能，支持对话闲聊、角色扮演、信息抽取等。
- 智能家居：为智能家居设备提供语音控制和交互功能，提升家庭智能化水平。
教育
在教育领域，MiniMax的大模型能够提供多种智能化服务：
- 课堂总结：自动生成课堂总结，帮助学生更好地理解和复习课程内容。
- 作文批改：智能批改学生作文，提供详细的反馈和建议。
- 英语口语练习：提供英语口语练习助手，帮助学生提升口语能力。
互动娱乐
MiniMax的大模型在互动娱乐领域也有广泛应用：
- 互动小说：用户可以在阅读小说的过程中与角色对话，提升阅读的沉浸感和趣味性。
- 虚拟角色：通过复刻知名IP，打造具有独特音色、说话语气和动作特征的虚拟角色，增强用户粘性和品牌影响力。
企业应用
MiniMax的大模型在企业应用中也展现了强大的能力：
- 办公协作：为办公软件提供文本生成、要点总结、任务分发等功能，提升办公效率。
- 客户服务：提供自动化客户服务，提升客户满意度和服务效率。
- 知识问答：为企业提供知识问答系统，帮助员工快速获取所需信息。
MiniMax的策略是部分开源、部分闭源。具体的开源与否取决于模型的类型和应用场景。对于一些基础工具和模型，MiniMax可能会选择开源以促进社区参与和技术创新；而对于核心的商业化产品，则主要采用闭源策略，以保护知识产权和商业利益。
MiniMax, founded in December 2021, quickly gained prominence in the AI field. The company independently developed three foundational models: text-to-vision, text-to-speech, and text-to-text, becoming the first startup in China to simultaneously possess large multimodal model capabilities across these three modalities.
ABAB Series Models
ABAB 5.5
- Features: ABAB 5.5 is a large language model developed by MiniMax, with an input-output limit of 8,192 tokens, suitable for character generation tasks.
- Performance: It performs well in writing, chatting, and Q&A scenarios, reaching the level of GPT-3.5. However, it has certain limitations in handling more complex tasks and those requiring precise outputs.
ABAB 6
- Features: ABAB 6 is MiniMax’s second version of the Mixture of Experts (MoE) large model, with a parameter count an order of magnitude larger than ABAB 5.5, better equipped to handle complex tasks.
- Technology: Utilizes the MoE architecture, where model parameters are divided into multiple groups of “experts,” with only a portion involved in each inference, improving computational efficiency and model performance.
- Performance: Performs excellently in testing across multiple complex tasks, significantly improving inference speed and training efficiency.
ABAB 6.5
- Features: ABAB 6.5 is a trillion-parameter large language model with a context window of 245k tokens, capable of handling complex productivity and multilingual dialogue scenarios.
- Performance: It approaches the leading industry models in tests involving knowledge, reasoning, mathematics, programming, and instruction-following.
Pricing Models
Token-Based Billing
- Text-to-Text (T2T): Suitable for interactive entertainment, text creation, and other scenarios. The billing unit is per 1,000 tokens, corresponding to about 750 characters (including punctuation), priced at 0.015 RMB per 1,000 tokens.
- Text-to-Speech (T2A): Suitable for speech broadcasting, voice interaction, and other scenarios. The billing unit is based on 10,000 characters (input), with one Chinese character counted as two characters, and English letters and punctuation as one character each.
Storage-Based Billing
- Vector Storage: Fees for vector storage retrieval are calculated daily as a weighted average, with a cost of 1.5 RMB per GB per day.
Session-Based Billing
- Code Interpreter: The billing unit is 0.2 RMB per session. A session is defined as one assistant calling the Code Interpreter in one thread within one hour. If it exceeds one hour, it is counted as two sessions.
Function-Based Billing
- Chat Completion Pro: This service offers more features than standard chat completion, including multi-person dialogue, conversation examples, plugins, and function calls. The price is 0.015 RMB per 1,000 tokens, and the web search function is priced at 0.03 RMB per search.
Free Trials
- Free Trial Credits: During internal testing, all accounts receive a fixed amount of free credits to experience API calls. Once live, users can recharge online and will be charged based on the number of tokens consumed.
Custom Services
- Model Fine-Tuning: Custom model fine-tuning services are offered, with pricing based on the number of tokens, the number of tokens in the input file, and the number of training epochs.
Text Processing
MiniMax’s large models excel in text processing and are suitable for the following scenarios:
- Logical Reasoning: Assists users with complex logical reasoning tasks and provides accurate reasoning results.
- Text Continuation: Automatically generates high-quality text continuations, suitable for novel and article creation.
- Copywriting Generation: Generates creative copy for advertising, marketing, and other scenarios.
- Content Summarization: Extracts key information from long texts, generating concise summaries.
- Code Generation: Generates code based on natural language descriptions, improving development efficiency.
Speech Synthesis
MiniMax’s speech models can generate natural, high-fidelity, personalized speech, applicable to the following scenarios:
- News Broadcasting: Generates professional news broadcast voices, enhancing information delivery.
- Voice Assistants: Provides voice interaction capabilities for smart hardware, improving user experience.
- Voiceovers and Audio Interaction: Delivers high-quality voiceover services for videos, games, and other media.
Smart Hardware
MiniMax’s large models have widespread applications in smart hardware:
- Mobile Voice Assistants: Provides smart voice assistant functions for mobile devices, supporting casual conversation, role-playing, and information extraction.
- Smart Home: Enables voice control and interaction for smart home devices, enhancing the level of home automation.
Education
In the education field, MiniMax’s large models offer various intelligent services:
- Class Summarization: Automatically generates class summaries, helping students better understand and review course content.
- Essay Grading: Intelligently grades student essays and provides detailed feedback and suggestions.
- English Speaking Practice: Provides an English speaking practice assistant, helping students improve their oral language skills.
Interactive Entertainment
MiniMax’s large models also have broad applications in interactive entertainment:
- Interactive Novels: Allows users to engage in dialogue with characters while reading novels, enhancing the immersion and enjoyment of the experience.
- Virtual Characters: Creates virtual characters with unique voices, speaking tones, and actions, replicating popular IPs to enhance user engagement and brand influence.
Enterprise Applications
MiniMax’s large models demonstrate strong capabilities in enterprise applications:
- Office Collaboration: Provides text generation, key point summarization, task distribution, and other functions for office software, improving work efficiency.
- Customer Service: Offers automated customer service, increasing customer satisfaction and service efficiency.
- Knowledge Q&A: Provides knowledge Q&A systems for enterprises, helping employees quickly access needed information.
Open-Source Strategy
MiniMax adopts a mixed strategy of partial open-source and partial closed-source. The decision to open-source depends on the type of model and application scenario. For some foundational tools and models, MiniMax may choose to open-source to foster community participation and innovation. For core commercial products, however, a closed-source approach is used to protect intellectual property and commercial interests.
