# Production inference. Recursive self-improvement.

Hedra is a research and infrastructure company bringing frontier inference to multimodal and physical AI. We serve your models and keep improving them.

[Get started](https://www.hedra.com/docs/pages/developer/v3/quickstart) · [Talk to an engineer](https://www.hedra.com/enterprise?form=true)

## Production Inference

### Your models, served faster.

Multimodal models push GPUs hard. We make yours faster with distillation and inference optimization, then run them on our infrastructure, so you don’t need your own serving stack.

[Run your model on Hedra](https://www.hedra.com/inference)

## Model Factory

### Models built for your problem.

General-purpose models fall short on real-world problems. Our forward deployed researchers join your team, learn the problem and your hardware, and build your model with you.

[Bring us your problem](https://www.hedra.com/model-factory)

## Recursive Self-Improvement · Research Preview

### Models that keep getting better.

Our researchers set the goals and guardrails. Agents run the experiments that keep improving your model. That is recursive self-improvement as a service.

[See how the loop works](https://www.hedra.com/research/rsi)

## Krea 3 inference optimization

Evidence · Krea 3 on AMD Instinct MI355X

22.9 s → 3.3 s

For Krea 3, our researchers tuned and then custom-implemented parts of the inference stack on AMD Instinct MI355X GPUs, using a quality check against a BF16 reference to decide which lower-precision settings to keep.

Median DiT time per edit fell from 22.9 s to 3.3 s for a 1 MP reference and 4 MP output on AMD Instinct MI355X GPUs, combined with model-level optimizations Krea provided. This measures DiT time, not full-request latency.

[Read the engineering write-up](https://www.hedra.com/blog/optimizing-krea-3-inference-on-amd-instinct-mi355x)

## Products

- [Inference](https://www.hedra.com/inference): Managed model APIs or optimized serving for your own model.
- [Model training](https://www.hedra.com/training): Post-train models with SFT and RL.
- [GPU Sandboxes](https://www.hedra.com/gpu-sandboxes): Isolated GPU environments for your own code.
- [World Model Runtime](https://www.hedra.com/wmr): An inference engine for image, video and world models.

## Applications

- [Physical AI](https://www.hedra.com/solutions/physical-ai)
- [Defense](https://www.hedra.com/solutions/defense)
- [Media generation](https://www.hedra.com/solutions/media-generation)

## Deployment

- [Deployment options](https://www.hedra.com/inference#deployment): Where models run, who operates them and what each option requires.
- [Sovereign deployment](https://www.hedra.com/sovereign): For workloads with jurisdiction, residency or control requirements.
- [Working with Hedra](https://www.hedra.com/enterprise#working-with-hedra): Deployment, security, ownership and commercial terms in one place.

## Research

- [Recursive self-improvement](https://www.hedra.com/research/rsi): What researchers decide, what agents run, and how accepted results change the next cycle.
- [Mechanistic interpretability](https://www.hedra.com/research/mechanistic-interpretability): How generative models represent distinct concepts, and how to steer one without disturbing the rest.
- [Post-training](https://www.hedra.com/research/post-training): Methods and evaluations for adapting multimodal models with SFT and reinforcement learning.
- [Inference](https://www.hedra.com/research/inference): Making large multimodal models fast and efficient on real hardware, with measured conditions.

## Bring us the problem you need to solve.

[Talk to our team](https://www.hedra.com/enterprise?form=true) · [Get started](https://www.hedra.com/developers)

Research expertise and GPU infrastructure for models you own.

[Company](https://www.hedra.com/company) · [Contact](https://www.hedra.com/contact) · [Privacy](https://www.hedra.com/privacy)