Hedra, a San Francisco-based artificial intelligence startup known primarily for video generation work, unveiled a new offering aimed at teams working in robotics, manufacturing, biology, and defence sectors. The new service, branded Hedra Model Factory, enables these organisations to develop custom AI models tailored to their specific operational needs, the company announced on Tuesday.
Rather than focusing on initial model training, the service concentrates on what happens after that foundational phase. Hedra supplies computational resources, assists in adapting models using proprietary customer data, manages model deployment, and handles ongoing refinement as systems operate in production environments.
The company combines human expertise with autonomous AI agents that conduct experiments independently. Data generated by deployed models feeds back into the development cycle, informing improvements to subsequent iterations.
The next step is the real world, where the data is video, sensors and simulation
Michael Lingelbach, Hedra's founder and chief executive
Organisations can supply their own simulation environments and datasets, with Hedra constructing training frameworks around these assets. For teams lacking internal research capabilities, Hedra personnel can manage data gathering and validation processes.
Deployment flexibility is built into the offering. Models may run on a customer's internal infrastructure or at a Hedra partner data centre facility. The company ensures that only model versions meeting customer-defined performance criteria reach production.
Real-world results
Hedra highlighted three use cases demonstrating the service's capabilities. First, the company optimised Krea 3, an image generation model from AI firm Krea, achieving a sixfold speed improvement in under seven days while maintaining the same AMD hardware.
In a second example, Hedra refined its own video generation model to better respect physical laws. The improved version achieved a Physics-IQ score of 34.9, a benchmark developed by Google DeepMind, compared to 30.3 for its earlier H3 Max iteration.
A third demonstration involved a simulated robotic system equipped with a new gripper mechanism. Without requiring a full retraining cycle, Hedra elevated the robot's task success rate from nearly zero to nearly 100%.
Growth and availability
The company reported that its annual revenue multiplied fifteenfold over the past year, though specific figures were not disclosed. Backing comes from venture capital firm Andreessen Horowitz.
Model Factory is now accessible through Hedra's website. Teams interested in the offering can also lease computing resources without committing to extended agreements, or execute existing models on Hedra's infrastructure.
Source: The Next Web
