Kernels built for vision
Fused operations for spatial-temporal attention, VAE encoding/decoding, and diffusion.
World Model
Runtime
Hedra’s open-core engine for image, video, world, and text models.
Optimized for diffusion workloads across NVIDIA and AMD GPUs.
Optimize the work of specialized models.
Diffusion steps, attention across frames, and VAE decoding dominate diffusion and world action model inference. WMR optimizes these operations together to reduce the GPU time each output needs.
Precision tested against quality
Quantization evaluated on image and video output quality.
Scheduling for diffusion
Batching and caching designed around fixed-step generation.
NVIDIA and AMD support
Run supported models across GPU vendors and hardware generations.
Run WMR on top GPUs.
Use the open engine with your own weights on supported NVIDIA and AMD GPUs, or get more performance via Hedra-hosted serving.
One stack for multimodal models.
VLAs and world action models
Run vision-language-action policies and action-conditioned world models for robotics and simulation.
Your models
Serve supported weights you own, license, or build with our team.
Video
Generate video from text or images, including streaming workloads.
Image
Generate, edit, and upscale images for interactive or batch workflows.
Text
Generate, annotate, and train text for top-tier LLM performance.
Audio
Serve speech, music, and sound models for voice and real-time applications.
Faster on the same GPU.
Time to generate one 5-second 768p clip
MiniMax H3 · One AMD MI355X
WMR Turbo
WMR
SGLang
2.6× faster
With WMR and auto optimization. 85% of clips accepted in blind quality review.
1.2× faster
MiniMax H3 on WMR vs. SGLang, with the same output.
Internal benchmark: MiniMax H3 · 5-second 768p clip · Single AMD MI355X · SGLang baseline · Aug–Sep 2026.
Use WMR through Hedra.
Dedicated inference
Serve your weights on a dedicated deployment.
GPU Sandboxes
Get direct GPU access to customize the runtime.
Sovereign Deployment
Hedra provides the serving stack and works with your operators.
Deploy with WMR.
WMR powers Hedra’s own models in production. Bring your workload to discuss performance and deployment.
