Report

Hedra puts AI on model-training duty with its new Model Factory

The new service combines AI agents, researchers and computing power to adapt models for robotics and other physical-AI workloads.

7 Sources, 1d ago, first seen 1d ago

Hedra's new Model Factory cranks up the production volume for physical AI model training - SiliconANGLE

TLDR

Hedra says its newly launched Model Factory pairs AI agents with researchers and computing power to adapt existing models for robotics, manufacturing, biology and defense. The proposed loop uses customer data, training experiments and feedback from deployed models. “Self-improvement” still has human supervisors, and the launch reporting doesn’t independently validate real-world gains.

Today's Rank

—

Not ranked yet

Combined views

—

7 Sources, first seen 1d ago

Bring a model and data

Hedra is targeting robotics, manufacturing, biology and defense workloads. Its offering centers on post-training: adapting an already-trained model for a particular task rather than building one from scratch. Source

According to CEO Michael Lingelbach’s description to SiliconANGLE, customers supply a base model and proprietary simulation and video data. Hedra provides a GPU sandbox, an isolated computing environment powered by graphics processors, while AI agents organize training runs, curate data and iterate against customer specifications. Researchers supervise the process. Lingelbach says vision-language models, which process images and text, score data and provide reinforcement-learning feedback: rewards that guide the model toward the desired behavior. Source

Hedra says customers can deploy the resulting models on their own premises or through its cloud data center partners. It also describes using distillation—training one model to imitate another—to prepare models for local devices that need fast responses. The report does not identify the partners, specify hardware requirements or give pricing. Source

After deployment, Hedra says models feed new data back into the customer’s sandbox to guide further training. Source

An image-model speedup

In one company-reported example covered by SiliconANGLE, Hedra used distillation to reduce a Krea 3 2K image-generation model’s latency from 22 seconds to 3.6 seconds. The account does not supply test-hardware details or a quality comparison. An image-generation speedup is also not a robot’s report card. Source

SiliconANGLE’s account includes no independent test showing that the proposed loop reliably improves performance on customers’ physical hardware. Source