MiniCPM-V2.6: MiniMax's Open-Sourced Multimodal Model for On-Device AI Gets an Upgrade

Tech 2024-08-14 00:49:33 Source: Network

MiniCPM-V2.6: MiniMax's Open-Sourced Multimodal Model for On-Device AI Gets an UpgradeOn August 7th, MiniMax, an artificial intelligence company, officially open-sourced its latest on-device AI multimodal model MiniCPM-V2

MiniCPM-V2.6: MiniMax's Open-Sourced Multimodal Model for On-Device AI Gets an Upgrade

On August 7th, MiniMax, an artificial intelligence company, officially open-sourced its latest on-device AI multimodal model MiniCPM-V2.6. This model, with a mere 8 billion parameters, achieves the best performance (SOTA) among models with less than 20 billion parameters in tasks such as single-image, multi-image, and video understanding.

The launch of MiniCPM-V2.6 marks a new milestone in the evolution of on-device AI multimodal models. This model brings several breakthroughs:

I. Feature Expansion:

Real-time Video Understanding: MiniCPM-V2.6 is the first to achieve real-time video understanding on edge devices. This means users can perceive and understand video content in real-time without relying on the cloud.

MiniCPM-V2.6: MiniMax

Multi-Image Joint Understanding: The model can process multiple images simultaneously and perform joint understanding and analysis, which is significant for applications like image search and multi-image scene comprehension.

Multi-Image ICL Visual Analogical Learning: MiniCPM-V2.6 enables cross-image analogical learning, transferring and sharing knowledge across multiple images. This equips the on-device model with stronger learning capabilities.

Multi-Image OCR: The model possesses the ability to recognize text content in multiple images, opening new possibilities for applications like automatic text recognition and information extraction.

II. Performance Breakthrough:

High Pixel Density: The pixel density of the MiniCPM-V2.6 model is twice that of GPT-4o, indicating it can process more information with the same memory footprint, thereby enhancing operational efficiency.

Fast Inference Speed: The quantized model requires only 6GB of memory and achieves an inference speed of 18 tokens per second on edge devices, 33% faster than its predecessor, ensuring rapid responses on edge devices.

Cross-Platform Support: The model supports multiple languages and inference frameworks, expanding its applicability and increasing its flexibility.

III. Upgraded OCR Capabilities:

Enhanced OCR Performance: MiniCPM-V2.6 maintains its SOTA performance in OCR while extending it to various scenarios such as single-image, multi-image, and video understanding.

Unified High-Definition Visual Architecture: The model utilizes a unified high-definition visual architecture to transfer and share OCR capabilities, enabling a smooth expansion from single-image to multi-image and video, significantly reducing the number of visual tokens and resource consumption.

Open Source and Accessibility:

MiniCPM-V2.6 is now open-sourced on GitHub and HuggingFace, allowing developers to freely access and utilize the model for further development and application.

The open-source release of MiniCPM-V2.6 not only invigorates the development of on-device AI multimodal models but also provides developers with more powerful and flexible tools, promoting the implementation of AI technology across diverse scenarios.

The release of this model will further fuel the advancement of on-device AI, delivering more convenient and intelligent experiences to users. As technology evolves, on-device AI will demonstrate its unique advantages in more domains, positively impacting humanity.

Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])

Mobile advertising space rental

Tag: MiniCPM-V2.6 MiniMax Open-Sourced Multimodal Model for On-Device AI Gets