Researchers and agents at work inside Hedra Model Factory

Designed for physical AI, Model Factory combines researchers and agents to adapt models and deploy them on the chosen infrastructure.

A research team supports each Hedra Model Factory project, from defining the desired behavior to deploying the model in production. The offering brings together post-training, experimentation, and inference optimization for teams working in robotics, manufacturing, biology, and defense. Customers provide their data and simulation environments; researchers can also help prepare them.

The claimed autonomy concerns the experiments assigned to agents. The official page leaves human researchers responsible for setting objectives and deciding which changes to retain. The service therefore combines research support with computing infrastructure, with the goal of carrying lessons from one project into subsequent cycles. An improvement loop that includes research tools

In the description of recursive self-improvement, experiments can target the model itself, agents’ research methods, inference software, or GPU execution. Each experiment must be assessed against the project’s objectives before its findings are carried forward. Improving the “factory” itself remains a research program whose priorities are outlined on the website.

To adapt models, the post-training research page describes two methods: supervised fine-tuning using examples, and reinforcement learning, in which outputs receive scores that guide updates. Evaluation must check the intended behavior and identify regressions in other capabilities. A higher reward score alone does not establish that a model is better for the application.

The GPU Sandboxes provide isolated environments where teams run their code, dependencies, and training loops. The pricing page still classifies training and these sandboxes as research previews. On 𝕏, the company describes infrastructure under development to support more than 100,000 simultaneous reinforcement learning environments and simulators. That figure is an announced target; no public capacity measurement was found in the pages reviewed. The Krea 3 case documents the tradeoffs

A detailed result appears in the engineering report on optimizing Krea 3 on AMD Instinct MI355X. Median execution time for the diffusion transformer, one component of the image-editing model, fell from 22.9 to 3.3 seconds. The comparison covers 23 edits supplied by Krea, using a 1 MP reference image and a 4 MP output. It does not measure the complete time between submitting a request and receiving the image.

The improvement combines model-level optimizations supplied by Krea with execution changes implemented by Hedra. Researchers worked on numerical precision, computational routines, and transfers between GPUs. Lower precision remained subject to quality checks: MXFP4 was rejected because it failed to meet the chosen threshold. Agents also tested approaches that produced no gains, including certain combinations of operations.

For World Models, the announcement thread on 𝕏 reports that retraining H3 produced a score of 34.9 on Physics-IQ. The official Physics-IQ repository describes an evaluation of physical understanding in video models. This reported result does not establish general robotics capabilities, and the protocol for this particular adaptation is not detailed in the documents located. Terms defined with each business

Deployment can take place on the provider’s infrastructure or the customer’s own systems. Production results are then intended to help determine which experiments to run next. The published pricing structure separates research and development, priced per engagement, from experimental compute and production inference. No single flat rate covers the entire Model Factory offering.

The promise of working on customer-owned models is further qualified in the engagement terms presented to enterprises. Rights to trained weights, checkpoints, and evaluation outputs are defined by agreement. That contract must also address data use and retention, base-model licenses, and export terms.