Hedra launches Model Factory for building and deploying AI models

Hedra describes Model Factory as a platform that brings together GPU sandboxes, post-training led by researchers and AI agents, low-latency deployment and a data flywheel for continued model improvement.[1] Hedra says its researchers work alongside autonomous AI agents during post-training.[1] The company is positioning the offering for teams in robotics, manufacturing, biology and defense.[1] Rather than presenting it as a single training tool, Hedra’s description connects compute, model development and deployment in one service.[1]

Customers bring their own data and simulation engines, while Hedra says it helps build reinforcement-learning environments and datasets.[1] For companies without a research lab, Hedra says its researchers can set up data capture, evaluations and the training-to-serving loop.[1] Hedra also says teams can scale training with GPU sandbox capacity without taking on long-term GPU commitments.[1]

Hedra says deployed models can run on a customer’s premises or at a nearby partner data center.[1] The company says it releases updated model versions only after they pass customer evaluations, and that each deployed model feeds data into the next version.[1] Hedra says it uses held-out cases to check whether training improved the target behavior while preserving capabilities the customer relies on.[3] These details describe an evaluation gate and a feedback loop; they do not establish how a particular customer’s model will perform.[1][3]

Hedra said Model Factory was available through its website starting on the day of the announcement, and its pricing page lists engagements as priced per engagement.[1][2] The page lists a rate of $0.001972 per GPU-second and AMD Instinct MI355X compute at $0.00081944 per GPU-second; it also labels training and GPU sandboxes as research previews.[2] Hedra reported cutting 2K image-editing time from about 22 seconds to 3.6 seconds in under a week, using the same four AMD GPUs for the comparison.[1] The comparison used a 30-step bf16 baseline and a distilled 14-step version.[1]

Hedra said a retrained H3 model scored 34.9 on Google DeepMind’s Physics-IQ benchmark, compared with 30.3 for H3 Max.[1] It also said an autonomously distilled H3 version rendered a five-second clip in under three seconds on AMD MI355X GPUs.[1] In a simulation study, Hedra reported restoring a robot’s task success from near zero to near 100% after a gripper change, without retraining.[1] These are results Hedra reported for specific tests; the evidence does not show whether the same outcomes will hold in other settings.[1]