Qwen2.5-VL-3B-Instruct

Qwen2.5-VL-3B is a compact yet powerful multimodal model with 3 billion parameters, specifically designed for efficient deployment on edge devices without compromising performance. The model combines a Vision Transformer encoder with the Qwen2.5 series decoder, enabling image, video, and text processing through innovative approaches like dynamic resolution handling and absolute temporal encoding. This allows it to process images of varying sizes and videos of up to one hour in length, with second-level event localization.

A key feature of the 3B model is its high-quality document analysis capability, including OCR, object detection, video understanding, and computer interface automation. Trained on approximately 4 trillion tokens across multiple modalities, the model achieves deep visual comprehension and can generate structured JSON outputs for object coordinates and attributes. The innovative Window Attention mechanism in its vision encoder significantly reduces computational costs, scaling linearly rather than quadratically—making it ideal for mobile and edge devices.

The model delivers outstanding results on key benchmarks: 53.1% on MMMU, 93.9% on DocVQA, 79.3% on TextVQA, and 62.3% on MathVista. Its performance in agent-based tasks is particularly impressive, scoring 76.9% on AITZ and 90.8% on AndroidWorld, demonstrating robust UI interaction capabilities.

Use cases for the 3B model span a wide range, from mobile document analysis (OCR) to security system integration. Licensed under Apache-2.0, it can be freely incorporated into commercial products, making it an attractive choice for startups and enterprises pursuing AI innovation.

Announce Date: 26.01.2025
Parameters: 4B
Context: 128K
Layers: 36
Attention Type: Full Attention
Developer: Qwen
Transformers Version: 4.41.2
License: Apache 2.0

Use our pre-built public endpoints for free to test inference and explore Qwen2.5-VL-3B-Instruct capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.