ORA: Smaller Models. Same Intelligence (www.oracomputing.com)

ORA has introduced an innovative model compression technology that significantly reduces the size of large language models (LLMs) while maintaining their intelligence and accuracy. This automated process enables models to shrink by up to 70%, allowing deployment on diverse hardware—from edge devices to cloud servers—within hours. The novel compression algorithm leverages information theory, surpassing traditional pruning and quantization methods, and claims minimal accuracy loss even at extreme compression levels. By accommodating popular models like Llama, Mistral, and Qwen, ORA’s solution is poised to impact various sectors, ensuring that enterprises can operate more efficiently and cost-effectively. The significance of ORA's announcement lies in its potential to drastically enhance AI model deployment across industries. With up to a 72% reduction in inference costs and a notable improvement in throughput (4.1 times higher performance on a single GPU), businesses can now run larger models without the burden of high computational costs. This advancement not only allows existing hardware to support more complex models but also offers a sustainable approach to AI development by reducing the overall memory footprint. As such, ORA's technology is set to empower silicon vendors, cloud providers, and enterprises alike, driving broader AI adoption while optimizing resource usage.