Hedra Launches Model Factory, Cuts a 2K Image Edit From 22 Seconds to 3.6

Hedra cut a 2K image edit from roughly 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs, and it is packaging the engineering behind that result as a product. The a16z-backed research company announced Model Factory on November 10, a platform that combines GPU infrastructure, model post-training (the process of adapting a pre-trained model to specific data and tasks), inference optimization, and production feedback into a single workflow.

The target customers are robotics, manufacturing, biology, and defense teams that need models adapted to their own data and operating conditions. The launch targets 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 can require custom training environments, specialized deployment software, and a way to evaluate each new version against the task it must perform.

What does Model Factory actually include?

The platform 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 (training setups where models learn by acting in simulated conditions and receiving feedback) 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 new versions only when they pass the customer's evaluations. That design 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. The announced process includes researchers and customer evaluation gates, rather than unrestricted automatic changes to production models.

Why physical AI needs more than a text-model serving stack

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

For a vision-language-action model (a model that produces actions as well as interpreting inputs), inference does more than generate text. In a simulated environment, the model and the environment generating its training experience can both require accelerators. Improving the workflow therefore involves more than selecting a model. It can mean reducing data movement, matching numerical precision to quality requirements, and coordinating training with serving.

The GPU sandbox offering 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 (small add-on modules that modify a model without full retraining)
• reinforcement learning, with checkpoint evaluation before serving

What does the sixfold speedup claim rest on?

The launch release highlights Krea 3 image editing as one example of the optimization process. Hedra says it reduced a 2K image edit from about 22 seconds to 3.6 seconds on the same four AMD Instinct MI355X GPUs in under a week.

The comparison combines inference optimization with distillation (training a smaller, faster model to imitate a larger one). The baseline used the original bf16 model at 30 sampling steps, while the faster version used a distilled 14-step model. That distinction matters. The result reflects changes to both the model's computation and the software executing it, and it is a reported result for a particular workload and configuration, not a general sixfold acceleration for every model the platform serves.

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. That write-up reports a 22.9-second-to-3.3-second comparison, but it measures diffusion-transformer time, not full-request latency. Those figures should be read separately from the launch release's image-editing result.

Video physics and robot hardware adaptation

Hedra also reports retraining H3, its video 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.

These examples point to different parts of the same engineering problem: preserving useful behavior while changing how a model runs or the system in which it operates.

A research service built around models customers own

Model Factory is available starting November 10. Teams can work with Hedra researchers on custom models, start with flexible GPU sandbox capacity without long-term commitments, or bring an existing model to its 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, although it did not disclose the revenue base in the announcement. Its company overview positions the platform as an extension of experience building and serving its own visual models.

For customers, the substantive proposition is a connected engineering loop: use domain data to improve a model, optimize it for the hardware and latency budget, evaluate the result, and feed production experience into the next iteration. How well that loop works will depend on task-specific evaluations and deployment constraints, as much as on the speed of the underlying GPUs.