OpenAI's Sora Video Model Expands to Microsoft Bing and Edge Devices
OpenAI's text-to-video model, Sora, is now being demoed on edge devices as part of a trend toward self-hosted, multimodal AI assistants. Microsoft has also integrated the technology into its search engine, launching free AI video generation powered by Sora directly within Bing.
- The version of Sora integrated into Bing, called Bing Video Creator, generates five-second-long, vertical-aspect-ratio videos at 480p resolution. While users receive a limited number of fast generations for free, further expedited video creation requires spending Microsoft Rewards points, which are earned by using Microsoft products like Bing and Edge. - Sora can generate videos up to 20 seconds long at a maximum resolution of 1080p for paid users of ChatGPT Pro. The model is a diffusion transformer, a technology that adapts the architecture used in image-generation models like DALL-E 3 to create video. - A significant challenge in deploying large generative models to edge devices is overcoming their memory and processing demands. This often requires model compression and quantization—techniques to reduce the model's size to run efficiently on the limited computational resources of embedded hardware. - Running AI models like Sora directly on edge devices, rather than in the cloud, is driven by the need for lower latency in real-time applications, enhanced data privacy by processing information locally, and improved reliability in environments with unstable internet connectivity. - Sora currently exhibits several key limitations, including a poor understanding of complex physics, cause-and-effect relationships, and spatial details like left and right. The model also struggles with object permanence, where elements in a scene may spontaneously deform or vanish during the video. - Microsoft's integration of Sora into its products is part of a multi-billion dollar partnership where Microsoft provides the Azure cloud infrastructure that powers OpenAI's research and model hosting. In turn, Microsoft integrates OpenAI's technology into services like Bing, Microsoft 365, and GitHub Copilot. - To ground Sora's creative outputs and ensure results are up-to-date, Microsoft employs an internal system called Prometheus, which blends the generative capabilities of OpenAI's models with factual data from its own Bing search index. - While heavily invested in OpenAI, Microsoft is also developing its own smaller, in-house AI models, such as Phi-4, to diversify its AI portfolio, reduce operational costs, and lessen its long-term reliance on a single partner.
