# Azion AI Inference Models import DocButton from '~/components/webkit/DocButton.vue'; Azion's edge-optimized models span multiple AI domains including text generation, image analysis, embeddings, and more. Each model is designed to balance performance and resource efficiency for edge deployment. This page provides a list of models available for use with **AI Inference**. To learn more about it, visit the [AI Inference Reference](/en/documentation/products/ai/ai-inference/). ## Available Models ### Mistral 3 Small (24B AWQ) This is a language model that delivers capabilities comparable to larger models while being compact. It is ideal for conversational agents, function calling, fine-tuning, and local inference with sensitive data. ### BAAI/bge-reranker-v2-m3 A lightweight reranker model with strong multilingual capabilities. It offers multilingual support and it's easy to deploy, with fast inference. ### InternVL3 InternVL3 is an advanced multimodal large language model with capabilities to encompass tool usage, GUI agents, industrial image analysis, 3D vision perception, and more. ### Qwen2.5 VL AWQ 3B A Vision Language Model (VLM) that offers advanced capabilities such as visual analysis, agentic reasoning, long video comprehension, visual localization, and structured output generation. ### Qwen2.5 VL AWQ 7B An instruction-tuned 30B-parameter FP8 causal language model for long-context (256K) text generation and reasoning, supporting chat/QA, summarization, multilingual tasks, math/science problem solving, coding, and tool-augmented workflows. ### Qwen3 30B A3B Instruct 2507 FP8 An instruction-tuned 30B-parameter FP8 causal language model for long-context (256K) text generation and reasoning, supporting chat/QA, summarization, multilingual tasks, math/science problem solving, coding, and tool-augmented workflows. ### Qwen3 Embedding 4B A 4B-parameter multilingual embedding model (36 layers, 32K context) that outputs 2560‑dim vectors for text/code retrieval, classification, clustering, and bitext mining. It supports instruction-conditioned embeddings and is optimized for efficient, cross-lingual representation learning. ### Nanonets-OCR-s An OCR model that converts document images to structured Markdown, preserving layout (headings, lists, tables) and basic tags. The output is easy to parse and feed into LLM pipelines. ### GPT-OSS 20B An OpenAI model with 20 billion parameters, designed for text generation, conversation, and various natural language processing tasks. This open-source model provides robust performance for a wide range of applications with tool calling capabilities and 131k token context length.