OVH announces Olala mini 7A1B (corporate.ovhcloud.com)

🤖 AI Summary
OVH has officially launched Olala mini 7A1B, its inaugural family of large language models (LLMs), set to be showcased at the OVHcloud AI Summit on November 19. This compact model features a 7-billion-parameter architecture with a unique hybrid attention mechanism, optimizing performance by only activating a portion of its parameters for each request. This innovative design allows it to operate efficiently on minimal hardware, running at 160 tokens per second on an RTX A4500 GPU and even on standard processors without a GPU. As an open-source model available on Hugging Face, Olala mini aims to enhance accessibility and foster innovation within the AI community. The development of Olala mini is significant as it showcases European capabilities in creating advanced LLMs, benefitting from public support and utilizing powerful supercomputers across Europe. The model has already seen practical application with Qwant, significantly improving response speeds and reducing operational costs. Future updates, including versions 1.1 and 2.0, aim to expand the context window and incorporate specialized features for various domains, positioning OVH as a strong player in the AI landscape. CEO Octave Klaba emphasizes the critical importance of developing proprietary models to reduce reliance on external resources, signifying a shift towards more localized AI solutions.
Loading comments...
loading comments...