Hedra launches Model Factory for specialist AI models

Hedra has launched Model Factory, a platform for building and running custom AI models for real-world applications. It is aimed at teams in robotics, manufacturing, biology and defence.

The platform combines access to GPU compute, post-training support, deployment infrastructure and a feedback loop designed to keep models improving after they enter production. Hedra is targeting customers that need models tailored to specific tasks and hardware, rather than relying on general-purpose systems.

That pitch reflects a growing commercial question in AI: how businesses will build and operate large numbers of specialist models for physical environments, where data often comes from video, sensors and simulations rather than text. In sectors such as manufacturing and robotics, companies can face high development and operating costs before systems are ready for deployment.

Hedra has opened the service to companies that want help creating models on their own data and running them in production. Its researchers work with customer teams on reinforcement learning environments, datasets, evaluations and the training-to-serving process.

How it works

The platform is structured around four stages: compute, post-training, deployment and self-improvement. Customers can use flexible GPU sandboxes on Hedra's inference engine rather than commit to long-term hardware contracts.

For post-training, teams can bring their own simulation tools and data while Hedra helps build the reinforcement learning environments and datasets needed for a given model. Its researchers also use AI tools that can caption and score data or provide feedback during online reinforcement learning.

Once models are ready, they can be run either on premises or at a partner data centre. The system is designed for low-latency inference and can also host existing customer models on Hedra's inference engine.

The final stage focuses on production monitoring and retraining. Hedra says researchers analyse how a model performs after deployment, retrain it and release new versions only if they pass customer evaluations.

Speed claims

Hedra used the launch to highlight internal research results intended to show the economics of optimisation. In one example, it said its team reduced the time needed for 2K image edits from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs.

According to Hedra, that improvement was achieved in under a week through model distillation and inference optimisation. The baseline used the original bf16 model at 30 sampling steps, while the faster version used a distilled 14-step model.

Hedra also pointed to work on physically consistent video models. It said a retrained version of its H3 model scored 34.9 on Google DeepMind's Physics-IQ benchmark, compared with 30.3 for H3 Max, while maintaining comparable visual quality. An autonomously distilled version rendered a five-second clip in under three seconds on AMD MI355X GPUs, according to the company.

In another example, Hedra described a simulation study involving robot hardware changes. It said a robot using a trained policy was fitted with a new gripper, and task success was restored from near zero to near 100% without retraining by matching the new gripper's reach and grasp point in the wrist camera view.

Business push

Hedra said revenue grew 15-fold over the past 12 months and that the new platform opens the research and infrastructure behind that expansion to outside teams. The company is backed by Andreessen Horowitz, known as a16z.

The launch places Hedra among a group of AI companies trying to commercialise tooling beyond foundation models and chatbots. Rather than selling a single general system, these groups are seeking revenue from adapting, optimising and maintaining models for narrowly defined industrial and physical tasks.

Michael Lingelbach, Founder and Chief Executive Officer of Hedra, said the company sees that as the next stage of AI deployment. "Recursive self-improvement is already changing how frontier models get built. The next step is the real world, where the data is video, sensors and simulation. Hedra Model Factory gives robotics, manufacturing, biology and defense teams the infrastructure to post-train, deploy and keep improving their own models," Lingelbach said.

Lingelbach has worked on reinforcement learning for visual and physical systems at Stanford and Nvidia. Hedra said its research team includes staff with experience at Stanford, Berkeley, UIUC, Nvidia, Meta Superintelligence Labs, Fireworks AI and Salesforce.

Customers can use the platform to build and deploy a custom model for tasks a general-purpose model cannot solve, or use Hedra's infrastructure to serve an existing model.