Baseten has demonstrated a new inference runtime that generates short videos with the open-source Wan 2.2 model in 2.75 seconds per clip. The company reports a 53.6-fold speed improvement over its baseline implementation, which took more than two minutes, and says the optimisation reduces its estimated generation cost from about five US cents to less than one-sixth of a cent per video.
The result is significant because video diffusion models have generally remained much slower and more expensive to serve than text models. Baseten is presenting the runtime as a demonstration of how model-level optimisation, low-level GPU engineering and scalable serving infrastructure can combine to move video generation closer to real-time use.
Three layers of optimisation
The largest speed gain comes from timestep distillation. Standard Wan 2.2 generation uses roughly 40 to 80 denoising steps. Baseten trained a distilled version that completes the process in four steps, reducing generation time from more than 120 seconds to about six seconds before the other runtime improvements are applied.
Baseten then optimised attention and other kernels used during inference. The company says its engineers produced a custom Triton attention kernel that was 1.58 times faster than the default implementation and fused several operations to reduce the overhead of repeatedly launching GPU kernels.
The third layer is NVFP4 quantisation on NVIDIA B200 GPUs. Moving selected operations from 16-bit to 4-bit precision increases available tensor-core throughput and reduces the cost of processing the model's linear layers. Baseten says the three techniques have multiplicative effects, producing the overall 53.6-fold improvement.
Infrastructure still matters
A fast model instance is only one part of a production video service. Baseten says its deployment also relies on rapid cold starts, autoscaling and request queues. Video generation typically uses a batch size of one, so horizontal scaling is necessary when traffic rises. Large container images and model weights can make cold starts slow, while queues must correctly route waiting requests to new replicas as capacity comes online.
The company is using its Baseten Delivery Network to reduce image and model-loading delays. Its demonstration separates queue time from generation time so users can see how both model execution and platform load affect the end-to-end experience.
A deliberate quality trade-off
The headline speed does not come without limitations. Baseten explicitly describes four-step timestep distillation as lossy. It says the model can be more likely to produce visual artefacts and may struggle with physical consistency in active scenes. That caveat matters when comparing the demo with slower workflows that use many more denoising steps.
The company also built a prompt classifier for the demonstration to reject unsuitable requests. Baseten says the classifier is configured to favour sensitivity, which can increase false positives. It recommends adapting safety classifiers to the expected prompt distribution when deploying a video model in a specific application.
Availability is currently limited
Baseten's public Wan 2.2 demonstration is available free until 31 July 2026. The announcement does not present the optimised runtime as a generally available, self-service production model API with a permanent price. Prospective customers should therefore distinguish the temporary demonstration from a contracted production deployment and confirm performance, output quality, safeguards and commercial terms for their own workload.
Even with that qualification, the work shows how quickly the economics of open video models can change when post-training, kernels, quantisation and serving infrastructure are treated as one system. For developers, the practical question is no longer only which model produces the best sample, but which combination of quality, latency, safety and cost is sustainable at production traffic.