Obsidian: Making Multi-Modal AI Accessible

Obsidian, an open-source artificial intelligence model, was announced on Reddit on October 30th, representing a breakthrough in the accessibility of multimodal AI. Obsidian is the first multimodal AI model with 3 billion parameters, making it compact enough to run efficiently on regular laptops.

Multimodal AI refers to AI systems that can process and connect data from different modalities, such as text, images, audio, and video. In this case, the model takes text and images as input, similar to OpenAI's latest version, GPT-4V. While impressive multimodal AI models like DALL-E 3 and GPT-4 demonstrate remarkable capabilities, their large size requires expensive high-end hardware to run, and their models are strictly confidential, making it impossible to run them even with the necessary dedicated hardware.

The Obsidian AI model fits multimodal intelligence into the memory of a standard laptop, achieving a breakthrough in the size of multimodal AI models. Built on the foundation of a 3 billion parameter Capybara-3B architecture, Obsidian achieves state-of-the-art performance compared to models of the same size. The developers also announced on Reddit that a multimodal model based on the acclaimed open-source 7B model, Mistral, will be released soon.

Obsidian's compact size is made possible by improved techniques from the LLaMA model architecture. According to the Reddit announcement, Obsidian was pretrained on a diverse synthetic multimodal dataset, which includes text and corresponding images. This training approach allows Obsidian to demonstrate powerful language and visual capabilities despite its reduced parameters.

The result is an AI assistant with conversational skills and visual understanding that can fit in your backpack. Obsidian breaks down barriers to accessing AI and opens up new possibilities for intelligence on devices.

Although it is currently in the early stages, Obsidian's efficient form factor presents an exciting future. It proves that multimodal AI doesn't have to be limited to massive data centers but can be made compact enough for widespread distribution.