Hedra Launches Model Factory to Accelerate Physical AI Model Training with Recursive Self-Improvement

Hedra Model Factory Brings Recursive Self-Improvement to Physical AI

Hedra, a San Francisco startup backed by Andreessen Horowitz (a16z), has launched Model Factory, a post-training and inference platform designed to accelerate AI model training for physical AI applications1

1

2

2

. The service targets robotics, manufacturing, biology, and defense teams, offering an autonomous training pipeline that uses recursive self-improvement (RSI) to continuously optimize custom AI models without the need for extensive retraining2

2

.

Michael Lingelbach, Hedra's founder and CEO, explained that physical AI requires specialized intelligence to handle real-world tasks, where data comes from video feeds, sensor readings, and 3D simulations2

2

. Traditional AI model training is not practical for physical AI because each robot and task requires a different model, making the compute costs prohibitively high2

2

.

How Model Factory Works with Autonomous Training Pipelines

Model Factory provides an end-to-end production platform where customers bring their base model and proprietary simulation or video data2

2

. Hedra provisions this data within a flexible GPU sandbox environment powered by its proprietary inference platform2

2

. The service pairs Hedra's in-house research teams with AI agents that orchestrate training runs, curate data, and iterate autonomously according to customer specifications2

2

.

Researchers use vision-language models to score data and provide live reinforcement learning feedback, continuously optimizing models for specific tasks2

2

. Once deployed—either on-premises or with Hedra's cloud data center partners—the models feed data back into the customer's sandboxed environment, creating an autonomous flywheel that improves performance over time1

1

2

2

.

Proven Results Across Robotics and Manufacturing Use Cases

Hedra demonstrated Model Factory's capabilities through several validation tests. The company retrained MiniMax Group's H3 video model to enhance physical consistency, improving its Physics-IQ benchmark score from 30.3 to 34.91

1

2

2

. In another case, Hedra accelerated the Krea 3 2K image generation model by six times, reducing latency from 22 seconds to 3.6 seconds while cutting sampling steps from 30 to 14—all achieved in under one week on AMD chips1

1

2

2

.

In a robotics demonstration, Hedra fixed a simulated robot that had been equipped with a new gripper, restoring its task success rate from near zero to near 100% without any retraining1

1

. The company achieved this by mapping the new hardware's reach and grasp metrics to the robot's wrist camera view, showcasing how RSI can bypass compute-intensive fine-tuning altogether2

2

.

Why Model Factory Matters for Physical AI Adoption

Lingelbach, a former reinforcement learning researcher at Stanford University and Nvidia, believes the next frontier for AI is the physical world, where data feedback loops from video, sensors, and simulations drive model improvement2

2

. RSI has already transformed how frontier models are developed in digital environments, but applying this continuous loop to physical AI has remained impractical due to massive multimodal data requirements2

2

.

Hedra's approach addresses this challenge by optimizing AI model training to make it cost-effective and scalable. The company reported 15 times revenue growth over the past year, though it did not disclose specific figures1

1

. Model Factory is now available on Hedra's website, with options for teams to rent computing power without long-term contracts or run existing models on Hedra's systems1

1

.

For enterprises, this service represents a way to bypass traditional training bottlenecks and quickly develop the specialized autonomous systems needed for real-world environments2

2

. Watch for how Hedra's platform influences adoption rates in sectors like defense and biology, where bespoke models are critical but historically expensive to develop.