Wan 2.2 / 2.6 / 2.7

On its owned track, Wan 2.2 (AX-WAN) generates short video clips from text or an image, with fine-grained control options — first/last-frame conditioning, pose or depth-guided motion, inpainting/outpainting within a frame, and audio-driven generation via speech-to-video. It's the model to reach for when cost efficiency matters more than chasing the newest generation's fidelity ceiling.

On its cloud track, Wan 2.6 and 2.7 offer the newer generation of the same architecture family as a hosted service, trading self-hosted cost savings for access to Alibaba's latest closed-weight improvements — useful when a shot needs more polish than the owned tier delivers and the higher per-clip cost is acceptable.

Wan is Alibaba's Apache 2.0-licensed video model family, and AutorunX runs it on two different tracks. The owned track is Wan 2.2 — a 14B mixture-of-experts (MoE) model alongside a smaller 5B variant — self-hosted on AutorunX's own Vast/RunPod GPU fleet under the internal name AX-WAN. Because the weights are openly licensed, AutorunX can run inference on its own infrastructure rather than paying a third-party API per call, which is what makes it the cost-efficient option in the lineup.

Wan 2.2's MoE architecture activates a subset of its parameters per inference step rather than the full network, which is part of how it keeps generation cost down relative to a dense model of similar total size. It supports a wide input surface: standard text-to-video and image-to-video, first-frame conditioning, first-and-last-frame (FLF2V) interpolation, VACE controls (depth maps, pose guidance, inpainting, and outpainting for targeted edits within a frame), and speech-to-video (S2V), where an audio track drives the generated motion. Clips run 5-10 seconds at 480p-1080p.

The cloud track — Wan 2.6 and 2.7 — is a separate, proprietary hosted tier reached through EvoLink rather than AutorunX's own hardware. Wan 2.6 supports image-to-video and text-to-video at 10-second clip lengths; Wan 2.7 adds a first_last_frame mode that takes two images (a start and end frame) to interpolate between. Both run at 720p-1080p. These are the same model family as the owned tier but a newer, closed-weight generation accessed as a paid API rather than self-hosted inference.