Hedra launches Model Factory to train custom AI and improve it after deployment

Hedra launched Model Factory on November 10, 2026, giving teams a service for adapting AI to their own data, deploying it and feeding production results into further training. Available immediately, it combines GPU computing with Hedra researchers and autonomous AI agents, targeting robotics, manufacturing, biology and defense rather than offering a single new general-purpose model.

Researchers stay inside the improvement loop

Customers bring their data and simulation engines. Hedra’s researchers help build the datasets and reinforcement learning environments—the settings where models practice tasks and receive feedback. For companies without a research lab, the service also covers data capture, evaluations and the connection between training a model and running it in an application.

The work includes AI tools that caption and score visual data or provide feedback during training. Hedra describes researchers working alongside autonomous AI scientists that run experiments and iterate on training runs. Its term for this process is recursive self-improvement: applying improvement methods to the way models themselves are developed.

That does not mean deployed models change without a release check. Once a model is running, Hedra researchers analyze production performance and retrain it. Replacement versions ship only after passing the customer’s evaluations. The announced workflow therefore keeps researchers and customer-defined tests between an experiment and a production update.

Three ways to start

• Custom-model work: collaborate with Hedra researchers to build and deploy a model around a specific problem.
• Compute capacity: use GPU sandboxes without long-term commitments. Teams can bring their own containers, dependencies and model code.
• Serving: bring an existing model to Hedra’s inference engine, the software that runs a trained model to produce outputs.

The speedup changed the model, too

Hedra’s clearest latency example is image editing. The company says it cut the time for a 2K edit in under a week, using the same four AMD Instinct MI355X GPUs. But it combined inference optimization with distillation, which changes the model rather than just the software executing it.

Hedra’s reported 2K image-editing result

Image-editing time

About 22 seconds→3.6 seconds

seconds

Company-reported comparison on identical hardware: the original model used 30 sampling steps; the distilled version used 14.

The original baseline used bf16 numerical precision and 30 sampling steps; the faster version used a distilled model with 14 steps. The roughly sixfold gain is consequently a result for that workload and configuration, not a platform-wide promise. It reflects less model computation as well as improvements in how that computation runs.

Physical behavior has its own tests

For video, Hedra reports retraining H3 for physical consistency. It scored 34.9 on Google DeepMind’s Physics-IQ benchmark, versus 30.3 for H3 Max, with what the company calls comparable visual quality. An autonomously distilled version rendered a five-second clip in under three seconds on MI355X GPUs.

A different example concerns robot hardware. In a simulation, Hedra replaced a robot’s gripper, then matched its reach and grasp point in the wrist-camera view. The company says task success recovered from near zero to near 100% without retraining. That finding applies to the simulation setup, not arbitrary hardware swaps in deployed robots.

Deployment can be on premises or at a nearby Hedra partner data center. The company describes optimization goals ranging from world-action models running faster than real time to vision-language-action models fitting on edge devices—equipment close to where a task happens. Those are deployment targets; the published performance examples remain specific company-reported results.