Hedra Launches Model Factory for Real-World AI Training

Hedra cut a 2K image edit from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs, and it did so in under a week. The a16z-backed research company is using that result as proof of what its new platform, Model Factory, can do. The product launches November 10 and targets robotics, manufacturing, biology, and defense teams that need models adapted to their own data and operating conditions.

Model Factory bundles four things into one workflow: GPU infrastructure, model post-training, inference optimization, and production feedback. The pitch addresses a practical gap in enterprise AI. A capable general-purpose model is often only the starting point. Turning it into a reliable system for video, sensor data, or physical actions requires custom training environments, specialized deployment software, and a way to evaluate each new version against its task.

How does the training loop work?

Model Factory combines flexible GPU sandboxes with post-training performed by Hedra researchers working alongside autonomous AI agents. Customers bring data and simulation engines. Hedra helps construct reinforcement learning environments and datasets, then optimizes how the resulting models run.

"Recursive self-improvement is already changing how frontier models get built," said Michael Lingelbach, founder and CEO of Hedra. "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."

After deployment, researchers analyze production performance, retrain the model, and release versions only when they pass the customer's evaluations. That makes evaluation criteria central to the service. A faster model still has to satisfy the requirements of the application using it. Hedra uses recursive self-improvement to describe researchers and autonomous AI scientists iterating on experiments, training runs, and optimization methods — a process that includes human oversight and evaluation gates, not unrestricted automatic changes to production models.

What sits underneath?

Hedra's World Model Runtime, or WMR, provides the inference foundation. The company describes it as an open-core inference engine for image, video, world, and text models, with support for NVIDIA and AMD GPUs. Its optimizations target diffusion steps, attention across frames, and encoding and decoding operations that consume substantial GPU time in visual workloads.

For a vision-language-action model, inference produces actions as well as interpreting inputs. In a simulated environment, both the model and the environment generating its training experience can require accelerators. Improving the workflow means more than selecting a model: it can involve reducing data movement, matching numerical precision to quality requirements, and coordinating training with serving.

The GPU sandbox lets teams bring containers, dependencies, and model code into isolated environments while Hedra manages placement, scaling, and hardware recovery. Teams can modify their runtime, run training or evaluation, and export checkpoints and artifacts. The training platform supports supervised fine-tuning, lightweight LoRA adapters, and reinforcement learning, with checkpoint evaluation before serving.

How solid are the speed claims?

The Krea 3 image-editing result deserves scrutiny. The baseline used the original bf16 model at 30 sampling steps. The faster version used a distilled 14-step model. The comparison combines inference optimization with distillation, so it reflects changes to both the model's computation and the software executing it. It is a reported result for one workload and configuration, not a general sixfold acceleration for every model served.

Hedra's separate Krea 3 engineering write-up details lower-precision matrix operations, custom attention, fused operations, and GPU-to-GPU transfers moved onto dedicated copy engines. Its 22.9-second-to-3.3-second comparison measures diffusion-transformer time, not full-request latency. Those figures should be read separately from the launch release's image-editing result.

What about video and robots?

Hedra also retrained its H3 model for physical consistency. The retrained model scored 34.9 on Google DeepMind's Physics-IQ benchmark, compared with 30.3 for H3 Max, with what Hedra describes as comparable visual quality. An autonomously distilled version rendered a five-second clip in under three seconds on MI355X GPUs, according to the company.

In another example, a simulation study replaced the gripper on a robot running a trained policy. Hedra says matching the replacement gripper's reach and grasp point in the wrist-camera view restored task success from near zero to near 100% without retraining. The finding concerns the reported simulation setup. It does not establish that arbitrary hardware changes can be handled without further training in deployed robots.

What does the business look like?

Model Factory customers can work with Hedra researchers on custom models, start with GPU sandbox capacity without long-term commitments, or bring an existing model to the inference engine. Deployment options include on-premises infrastructure or a nearby partner data center. Optimization targets range from faster-than-real-time world action models to vision-language-action models that fit on edge devices.

The company says revenue grew 15 times over the preceding 12 months, though it did not disclose the revenue base. The substantive proposition is a connected engineering loop: improve a model with domain data, optimize it for the hardware and latency budget, evaluate the result, and feed production experience into the next iteration. How well that works will depend on task-specific evaluations and deployment constraints as much as on GPU speed.