Catalog

Models that are ready for production.

Windflow takes the best open-source video and world models, then does the unglamorous work — kernels, schedulers, quantization, eval — to make them usable behind a real product. Every model in the catalog is benchmarked daily against the same SLO.

Model

Family

Resolution

FPS

p50 latency

License

Helios
Interactive real-time video generation with infinite streaming. Persistent geometry, physics and lighting across thousands of frames — built for game and simulation workloads.
Real-time long video
1280×720
24
47 ms
Apache 2.0 (PKU-YuanGroup)
Interactive real-time video generation with infinite streaming. Persistent geometry, physics and lighting across thousands of frames — built for game and simulation workloads.
LingBot
High-fidelity environments with action-conditioned interactive output. Steering, control input and gameplay signals are fed back into every step.
Action-controlled world model
1280×720
16
52 ms
Apache 2.0 (Robbyant)
High-fidelity environments with action-conditioned interactive output. Steering, control input and gameplay signals are fed back into every step.
LongLive
Multi-shot interactive video generation. Hold narrative state — characters, props, lighting — across scene cuts while the user keeps steering.
Multi-shot video
1024×576
24
58 ms
Apache 2.0 (NVIDIA / MIT)
Multi-shot interactive video generation. Hold narrative state — characters, props, lighting — across scene cuts while the user keeps steering.
CausVid
Autoregressive video diffusion: low-latency text-to-video and frame-accurate video-to-video, including restyle, background replace and virtual production composition.
Causal video diffusion
1280×720
30
38 ms
Research-only (MIT / Adobe), commercial via Windflow
Autoregressive video diffusion: low-latency text-to-video and frame-accurate video-to-video, including restyle, background replace and virtual production composition.
LTX-Video
Lightweight streaming diffusion tuned for mobile and edge devices. Runs inside a 6 GB working set with no quality cliff.
Real-time DiT video
896×504
20
55 ms
Apache 2.0 code, OpenRAIL-M weights (Lightricks)
Lightweight streaming diffusion tuned for mobile and edge devices. Runs inside a 6 GB working set with no quality cliff.
HunyuanVideo
Large open-weights text-to-video model, re-targeted by Windflow for streaming output and mid-generation prompt edits.
Text-to-video
1280×720
24
62 ms
Tencent Hunyuan Community License (not available in EU / UK / KR)
Large open-weights text-to-video model, re-targeted by Windflow for streaming output and mid-generation prompt edits.