Skip to content

Text-To-Text

MaaS-Llama

The MaaS-Llama series of models have high performance and a wide range of application prospects.

The following models are available for purchase:

  • MaaS Llama 3.1 70B Instruct
  • MaaS Llama 3.1 8B Instruct

MaaS Llama 3.1 70B Instruct

MaaS Llama 3.1 70B Instruct is a high-level language model with 70 billion parameters, specifically designed and optimized for instruction-following tasks. This model is based on the latest deep learning techniques and can understand and execute complex instructions, suitable for various natural language processing tasks such as text generation and Q&A.

  • Parameter Count: With 70 billion parameters, it can abstract and generalize complex language patterns and contexts.

  • Instruction Optimization: Specifically optimized for instruction-following tasks, it can accurately execute user instructions.

  • Multi-task Adaptability: Suitable for various natural language processing tasks such as text generation, Q&A, and dialogue.

  • High Accuracy and Reliability: Achieves high accuracy and reliability through large-scale pre-training, capable of providing intelligent language services.

MaaS Llama 3.1 8B Instruct

MaaS Llama 3.1 8B Instruct is a high-level language model with 8 billion parameters, specifically designed and optimized for instruction-following tasks. This model is based on the latest deep learning techniques and can understand and execute complex instructions, suitable for various natural language processing tasks such as text generation and Q&A.

  • Parameter Count: With 8 billion parameters, it guarantees performance while optimizing resource use.

  • Instruction Optimization: Specifically optimized for instruction-following tasks, it can accurately understand and execute instructions.

  • Multi-task Adaptability: Suitable for various natural language processing tasks such as text generation and Q&A.

  • High Efficiency and Reliability: Trained on large-scale data, it possesses high efficiency and reliability, capable of providing intelligent language services.

MaaS-C

MaaS-C is a robust natural language processing model, distinguished by its powerful capabilities in language comprehension and generation. It adeptly understands complex semantic relationships and contextual information, producing high-quality, fluent, and natural text.

The following models are now available for purchase:

  • MaaS-3.5 Sonnet
  • MaaS-3 Sonnet
  • MaaS-3 Haiku （On demand）

MaaS-3.5 Sonnet

MaaS-3.5 Sonnet is the inaugural version of the MaaS 3.5 series, boasting enhanced speed and superior capabilities in coding, visual interpretation, and natural language understanding.

  • Multifarious Capabilities Surpassing Predecessors

    In various performance tests across reading, programming, mathematics, and visual processing, it demonstrates exceptional proficiency. It shows marked improvement in understanding subtle nuances, humor, and complex instructions, with the ability to create high-quality content in a natural and appropriate tone.

  • Potent Visual Capabilities

    Excelling in tasks involving the interpretation and analysis of visual data, it comprehends complex charts, graphs, and diagrams, analyzes infographics and scientific visualizations, and explains spatial relationships and contexts within scenes. It can seamlessly integrate image and textual information, accurately recognize and describe objects within images, perform visual question answering, and leverage visual data to aid in problem-solving, such as analyzing architectural plans or engineering diagrams. Additionally, it offers insights in art and design analysis, exhibits improved handwriting text recognition from imperfect images, processes various text styles and languages, comprehends the context of text within images, and often retains or describes the original formatting when transcribing structured text.

  • Wide Range of Applications

    It can be employed in customer service, content creation, educational tutoring, programming assistance, data analysis, and other scenarios, potentially giving rise to entirely new business models and services.

MaaS-3 Sonnet

  • Multilingual and Multitasking Capabilities

    Exceling in analysis, prediction, meticulous content creation, code generation, and multilingual conversations, it is capable of summarizing approximately 150,000 words, handling multiple languages, and executing various tasks such as comprehending complex texts and coding.

  • Visual Processing Abilities

    It can manage photos, charts, graphs, and technical diagrams, as well as formatted materials like PDFs, flowcharts, or slides. It accurately identifies and describes objects within images, performs visual question answering, and utilizes visual information to assist in problem-solving, such as analyzing architectural plans or engineering diagrams, and providing insights in art and design analysis.

  • Long-Text Processing Capabilities

    The model can accept texts up to 200,000 words and has the capacity to remember over 1 million tokens.

  • Faster Execution Speed

    With enhanced execution speed, it can handle tasks and respond to user requests more efficiently.

MaaS-3 Haiku

  • Rapid Speed

    Capable of reading a paper with dense charts and graphical information (approximately 10,000 tokens) in under three seconds.

  • High Cost-Performance Ratio

    Offers significant cost-effectiveness in the intelligent category market.

  • Enhanced Multilingual and Multitasking Abilities

    Improved capabilities in analysis, prediction, content creation, code generation, and multilingual dialogues.

  • Visual Processing Abilities

    Adept at handling various visual formats, including photos, charts, graphs, and technical illustrations. It can accurately identify and describe objects within images, perform visual question answering, and leverage visual information to assist in problem-solving.

  • Long-Text Processing Abilities

    Supports a context window of 200,000 tokens upon launch and can accept over 1 million tokens as input, with a strong memory capacity to accurately recall information from vast amounts of data.

MaaS-Ge

The MaaS-Ge model is a high-performance, multitasking AI model, renowned for its exceptional precision and efficiency. It adeptly handles a diverse range of tasks, showcasing remarkable adaptability and flexibility. Moreover, the MaaS-Ge model is designed with scalability in mind, allowing it to be effortlessly deployed and optimized across various application scenarios to meet diverse business demands.

The following models are now available for purchase:

  • MaaS-1.0 Pro （On demand）
  • MaaS-1.5 Pro
  • MaaS-1.5 Flash （On demand）

MaaS-1.0 Pro

The MaaS-1.0 Pro exhibits high flexibility and adaptability, excelling in specific fields and application scenarios.

  • Multimodal Capabilities

    Capable of processing inputs from text, images, audio, and video, and generating text and image outputs. For instance, in educational contexts, it can comprehend physics problems, accurately recognize messy handwritten content, understand the structure of questions, convert problems and solutions into mathematical typesetting, identify erroneous steps in students' problem-solving processes, and then provide correct solutions.

  • Long Context Support

    Supports context lengths of up to 32k, enabling better handling of lengthy texts and complex task scenarios.

  • Efficient Attention Mechanisms

    Employs efficient attention mechanisms, such as multi-query attention, to enhance the model's processing efficiency and accuracy.

MaaS-1.5 Pro

  • Exceptional Contextual Processing Capacity

    Capable of handling information containing up to 1 million tokens, it can comprehend extensive documents of up to 1500 pages in one go, summarize 100 emails, process an hour-long video, or manage a codebase exceeding 30,000 lines.

  • Multimodal Input Support

    Proficient in simultaneously processing and understanding text, image, video, and audio data, it excels in handling complex scenarios rich in information, such as video content comprehension and multi-language translation tasks.

  • Outstanding Performance in Complex Tasks

    Displays significant advancements in handling complex prompts and coding tasks, better addressing challenging task scenarios.

  • Efficient Reasoning Capabilities

    Through innovative architecture and training methodologies, it accurately recalls and infers detailed information from extensive contextual data.

  • Promoting Multi-Field Application Development

    Paves the way for breakthroughs in long-document Q&A, long-video Q&A, and long-context automatic speech recognition, while also providing robust support for practical applications in education, research, media, and numerous other fields.

MaaS-1.5 Flash

MaaS-1.5 Flash is a lightweight and efficient AI model with multimodal processing capabilities, capable of handling text, images, audio, and video simultaneously.

  • High-Speed Response

    The optimized architecture enables rapid response when processing large volumes of data, generating output at the fastest speed across all test languages, ensuring users receive answers and results more quickly, thereby enhancing interaction and usage efficiency.

  • Extended Context Window

    Supports contextual processing of up to two million tokens, allowing it to tackle complex tasks such as long video analysis and multi-chapter document generation. For example, it can comprehend multiple extensive documents totaling up to 1500 pages in a single go, summarize 100 emails, process an hour-long video, or manage a codebase exceeding 30,000 lines.

  • Multimodal Processing

    Not only able to handle text, but it also delves deeply into analyzing and understanding images, audio, and video. For instance, it can identify the main content and visual details in videos, analyze and answer questions from image and video inputs, although it still has some shortcomings when processing audio inputs.

  • Versatile Task Application

    Excels in tasks such as summarization, chat applications, providing image captions and video subtitles, and extracting data from long documents and tables.

MaaS-GP系列

MaaS-o1

MaaS o1 model is specifically designed to handle reasoning and problem-solving tasks, with improved specificity and functionality. These models spend more time processing and understanding user requests, and they exhibit exceptional strength in fields such as science, coding, and mathematics compared to their earlier iterations.

Version Description Maximum Request/Tokens Support Status
MaaS o1 preview 2024-09-12 The o1 series features the most powerful model, offering enhanced reasoning capabilities. Input: 128,000
Output: 32,768		 Supported
MaaS o1 mini 2024-09-12 The o1 series includes faster and more cost-effective options, making them ideal for coding tasks that require speed and lower resource consumption. Input: 128,000
Output: 65,536		 Supported

MaaS-4o

MaaS-4o integrates text and images within a single model, allowing it to simultaneously process multiple data types. This multimodal approach enhances the accuracy and responsiveness of human-computer interactions. Comparable to MaaS-4 Turbo in English text and coding tasks, MaaS-4o surpasses it in performance for non-English languages and visual tasks, setting new benchmarks for AI capabilities.

Version Description Maximum Request/Tokens Support Status
Maas-4.0 (2024-08-06) Maas-4.0 (2024-08-06) contains all the features of the previous version, as well as:
1. Enhanced functionality for structured output extraction.
2. The maximum output token count has increased from 4,096 to 16,384.		 Input: 128,000
Output: 16,384		 Supported
MaaS-4o mini (2024-07-18) The latest compact GA model
1. A fast, affordable, and powerful model, an ideal replacement for the MaaS 3.5 Turbo series.
2. Text and image processing.
3. JSON mode.
4. Parallel function calling.
5. Enhanced features not supported.		 Input: 128,000
Output: 16,384		 Supported
MaaS-4o (2024-05-13) The latest large GA model
1. Text and image processing.
2. JSON mode.
3. Parallel function calling.
4. Enhanced accuracy and responsiveness.
5. Comparable to the visual-enabled MaaS-4 Turbo for English text and coding tasks.
6. Superior performance in non-English languages and visual tasks.
7. Enhanced features not supported.		 Input: 128,000
Output: 4,096		 Supported

MaaS-4 Turbo

MaaS-4 Turbo is a large multimodal model (accepting both text and image inputs to generate text), optimized for chat functionality similarly to MaaS-3.5 Turbo and the earlier MaaS-4 models. It excels in handling conventional completion tasks proficiently.

Version Description Maximum Request/Tokens Support Status
MaaS-4 Turbo (2024-04-09) The latest GA model
1. A replacement for all MaaS-4 preview models (vision-preview, 1106-Preview, 0125-Preview).
2. Feature availability currently varies based on input method and deployment type.
3. Enhanced features not supported.		 Input: 128,000
Output: 4,096		 Supported

It is a replacement for the following preview models:

  • MaaS-4 version：1106-Preview
  • MaaS-4 version：0125-Preview
  • MaaS-4 version：vision-preview

MaaS-4

MaaS-4 is the predecessor of MaaS-4 Turbo. Both the MaaS-4 model and the MaaS-4 Turbo model share the foundational model name MaaS-4. The distinction between the MaaS-4 model and the Turbo model can be made by examining the model version.

Version Description Maximum Request/Tokens Support Status
MaaS-4 (0125-Preview)
Preview Version of MaaS-4 Turbo		 Preview Model
1. Replaced 1106-Preview
2. Enhanced code generation performance
3. Reduced instances of incomplete tasks
4. JSON mode
5. Parallel function calls
6. Reproducible outputs (preview)		 Input: 128,000
Output: 4,096		 Supported
MaaS-4 (vision-preview)
MaaS-4 Turbo with Vision Capabilities Preview		 Preview Model
1. Accepts text and image inputs
2. Supports enhanced features
3. JSON mode
4. Parallel function calls
5. Reproducible outputs (preview)		 Input: 128,000
Output: 4,096		 Supported
MaaS-4 (1106-Preview)
MaaS-4 Turbo Preview Version		 Preview Model
1. JSON mode
2. Parallel function calls
3. Reproducible outputs (preview)		 Input: 128,000
Output: 4,096		 Supported
MaaS-4-32k (0613) Older GA Model
1. Basic function call using tools		 32,768 Upon Request
MaaS-4 (0613) Older GA Model
1. Basic function call using tools		 8,192 Upon Request
MaaS-4-32k (0314) Older GA Model 32,768 Upon Request
MaaS-4 (0314) Older GA Model 8,192 Upon Request
  • Compared to MaaS-4-1106-preview, MaaS-4 version 0125-preview more thoroughly accomplishes tasks such as code generation. Hence, depending on the task, clients might find that MaaS-4-0125-preview produces more output than MaaS-4-1106-preview. We recommend that clients compare the outputs of the new model. MaaS-4-0125-preview also addresses a bug in the UTF-8 handling for non-English languages that was present in MaaS-4-1106-preview. MaaS-4 version turbo-2024-04-09 is the latest GA version, superseding 0125-Preview, 1106-preview, and vision-preview.

MaaS-3.5

The MaaS-3.5 model is capable of understanding and generating natural language or code. The most powerful and cost-effective model in the MaaS-3.5 series is MaaS-3.5 Turbo, which has been optimized for chat and excels at traditional completion tasks. MaaS-3.5 Turbo is available for use with the chat completion API. The MaaS-3.5 Turbo instructions provide functionality similar to text-davinci-003 when using the completion API rather than the chat completion API.

Version Description Maximum Request/Tokens Support Status
MaaS-3.5-turbo-0125 Latest GA Model
1. JSON Mode
2. Parallel Function Calls
3. Reproducible Outputs (Preview)
4. Higher Accuracy in Responding in Requested Format
5. Bug Fixes for Text Encoding Issues in Non-English Function Calls		 Input: 16,385
Output: 4,096		 Supported
MaaS-35-turbo (1106) Previous GA Model
1. JSON Mode
2. Parallel Function Calls
3. Reproducible Outputs (Preview)		 Input: 16,385
Output: 4,096		 Available on Request
MaaS-35-turbo-instruct (0914) Completion Endpoint Only 4,097 Supported
MaaS-35-turbo-16k (0613) Previous GA Model
1. Basic Function Calls Using Tools		 Input: 16,384 Available on Request
MaaS-35-turbo (0613) Previous GA Model
1. Basic Function Calls Using Tools		 Input: 4,096 Available on Request
MaaS-5-turbo (0301) Previous GA Model Input: 4,096 Available on Request