Inference

Run your multimodal model in production on Hedra or your infrastructure. We tune the model and serving stack together for your workload and GPUs.

Talk to our teamDeployment options ↓

From model to production.

Find the bottleneck

We profile model execution and data transfers between GPUs to see where time goes.

Tune the model and runtime

We test distillation, kernel tuning, lower precision or GPU scheduling based on what limits performance.

Compare to the baseline

We measure latency, throughput and output quality before choosing what to serve.

Explore World Model Runtime

Krea 3 on AMD Instinct MI355X

22.9 s → 3.3 s median DiT time per edit

We tuned numerical precision, GPU kernels and communication for Krea 3’s image-editing workload. Our optimized stack, combined with model-level optimizations Krea provided, reduced diffusion transformer (DiT) time per edit.

Medians across 23 image edits after warm-up, with a 1 MP reference and 4 MP output. Measures DiT time, not full-request latency.

How we optimized Krea 3 on AMD GPUs

Deployment options

Hedra Cloud

Where it runs
On GPU capacity Hedra provides.
Who operates it
Hedra runs the stack.
Discuss your model

Your cloud

Where it runs
In your cloud account. We agree on the hardware and configuration with you.
Who operates it
We agree with your team on who runs each part of the stack.
Discuss deployment

Sovereign deployment

Where it runs
On infrastructure you control, within agreed data boundaries.
Who operates it
Your team runs the stack with support from Hedra.
Explore sovereign deployment
Security, ownership and engagement terms

Bring us your model.

Talk to our team