Hedra

Hedra is a research and infrastructure company that serves multimodal and physical AI models and keeps improving them. For production inference, it optimizes a customer's model with distillation and inference optimization, then runs it on Hedra's infrastructure, so the customer does not need its own serving stack.

Model Factory is for teams that need a model built for their problem. Forward deployed researchers join the team, learn the problem and the hardware, and build the model with them. An engagement does not require a finished model or a complete dataset to start.

The product range includes Inference, which offers managed model APIs or optimized serving for your own model; Model training, which post-trains models with SFT and RL; GPU Sandboxes, which are isolated GPU environments for your own code; and World Model Runtime, an inference engine for image, video and world models. The company names physical AI, defense and media generation as application areas.

Hedra also runs a research preview called recursive self-improvement. Its researchers set the goals and guardrails, and agents run the experiments that keep improving a model. On a Krea 3 project, the company reports that median DiT time per edit fell from 22.9 s to 3.3 s on AMD Instinct MI355X GPUs. The page notes that this measures DiT time, not full-request latency.

GPU compute is priced per GPU-second. Listed rates run from $0.000842 per GPU-second for NVIDIA RTX PRO 6000 to $0.001972 for NVIDIA B300. Model Factory uses custom pricing per engagement, and reserved capacity requires a quote. Sovereign deployment is available for workloads with jurisdiction, residency or control requirements.