Alibaba Cloud provides a group of technical pillars as the foundation of your multimodal AI development:
Qwen-Audio: The large audio language model that processes various audio inputs.
Qwen-VL: The large vision language model that analyzes images and reveals nuanced details and text.
OpenSearch: The conversational search service that tailors your Q&A system, leveraging vector retrieval and LLMs.
Qwen-Agent: The framework for developing LLM applications. It orchestrates intelligent agents that follow instructions and execute complex tasks.
PAI-EAS: You can deploy Qwen and other popular multimodal AI models as web services with just a few clicks.

These products are hinged on each other by a planner agent; you can deploy the API of this agent on ECS, and connect to DingTalk or other instant messengers (IMs).

PAI-EAS (Elastic Algorithm Service) is a component of Machine Learning Platform for AI (PAI) that allows users to deploy and manage AI applications easily. It includes the PAI-Blade inference accelerator, a hardware accelerator engineered to improve the performance of LLMs with an inference speed up to six times faster than traditional methods. This means you can process more data, generate faster responses, and enhance the performance of your language-based applications. You can run Llama 2 on PAI-EAS in WebUI with just a few clicks and the sample code provided in the blog, or use API with Python Wrapper code in the blog to set up Llama 2 service. You can also find guidance on how to mount a custom model on Llama 2 with PAI-EAS in the blog.

You can use PAI-DSW (Data Science Workshop) to fine-tune the Stable Diffusion models. Utilize PAI-DLC (Deep Learning Container) to quickly set up and manage a streamlined deep learning environment with ready-to-use containers of popular deep learning frameworks. Finally, deploy the models and create text-to-image generation service with a few clicks in PAI-EAS.

AnalyticDB for PostgreSQL is a cloud-native data warehouse with massive parallel processing (MPP) capabilities. The vector database feature can help you manage large-scale unstructured data such as documents, emails, image and web content and is ideal for modern data services that manage and process large volumes of structured and unstructured data, provide real-time analysis, and adapt to changing business needs. Enterprises with domain-specific requirements can leverage AnalyticDB for PostgreSQL to produce accurate results on topics based on domain-specific or specialized data. In this example, you can easily build a chatbot that combines LLM’s learning and comprehension capabilities with the core capabilities of AnalyticDB for PostgreSQL. This includes built-in vector data retrieval and full-text retrieval, providing professional and timely answers with the knowledge of your enterprise.

OpenSearch provides a high-performance search service that supports billion-level 128-dimensional documents (as vector data), or thousand-level QPS with millisecond-level query response. It integrates different retrieval algorithms for vector data and supports complex search requirements such as hybrid search, filtering by expression, and filtering while searching. OpenSearch also provides data compression features and fine index structure design to lower storage and memory requirements, saving hardware costs at scale. You can combine OpenSearch with LLMs through a service tool, to build an intelligent, dedicated, and real-time search service for your business. This will improve service quality and timeliness, without requiring heavy investment in hardware.

LoRA (Low-Rank Adaptive Relational Attention) is a widely used algorithm for AI image generation which allows users to fine-tune generative models by adding a small number of parameters, and using a small number of data sets. Platform for AI (PAI) allows you to easily and quickly fine-tune AI generated imagery based on LoRA and Stable Diffusion models, providing a wide range of generation space for models, poses, and backgrounds while reducing the number of trainable parameters and GPU memory requirements. This opens up new possibilities for a more personalized and intelligent customer experience for the fashion retail industry.

EasyDispatch is an AI-powered logistics management platform that enables businesses to optimize their logistics operations through real-time monitoring, intelligent dispatching, and predictive maintenance capabilities. The platform uses advanced algorithms to optimize delivery routes, minimize transportation costs, and improve overall efficiency. You can combine Alibaba Cloud AnalyticDB and a LLM to build a chat-style plugin for EasyDispatch to create and manage delivery tasks without the need for manual operations. This plugin can provide personalized and accurate responses to customer inquiries, and optimize delivery routes automatically to improve efficiency.