Hedra

Hedra is a visual AI research lab that sells access to its model catalogue, fine-tuning recipes, GPU sandboxes, and the WMR inference engine, deployable as managed cloud, hybrid, or sovereign on-prem infrastructure for teams working with diffusion, video, and world action models.

Hedra is a research and infrastructure company rather than a single consumer app. The lab builds frontier visual models, publishes work on video clip generation and simulation, and packages the same engine it uses internally as a platform for outside teams. That engine, WMR (World Model Runtime), is described as purpose built for diffusion, video, and world action models. The premise is straightforward: the compute machinery that serves Hedra's own multimodal workloads is available to customers who bring their own models or data. The product breaks into three layers. Managed Inference gives access to a catalogue of more than 100 models over an API, with Hedra handling hosting and scale. Pre and Post-Training lets teams fine-tune supported models on their own data using Hedra's training recipes, evaluate the results, and deploy on the same stack, including reinforcement learning for task-specific goals. GPU Sandboxes provide isolated environments created through the Hedra SDK, where custom model code, video pipelines, and simulations run with self-managed dependencies while Hedra handles GPU scheduling and scaling. Technical differentiation sits in how the runtime treats visual workloads: vision-native kernels that optimize attention across frames and image encoding and decoding, quality-aware quantization that trades numerical precision while watching image and video quality, and step-aware scheduling that arranges batching and caching around generation steps to improve latency and throughput. Deployment comes in three shapes: Hedra Cloud with the infrastructure operated for you, hybrid setups that place WMR and Hedra building blocks inside your own environment, and sovereign on-prem clusters under customer or in-country control. This is aimed at machine learning engineers, research teams, and media or robotics organizations with visual model workloads and enough technical depth to work from an SDK and API. Anyone looking for a point-and-click creative studio will find the site oriented toward developers and procurement conversations instead.