What Wan 2.7 is good at — and what it's not

Wan 2.7 is the model that broke the closed-source moat on flagship-tier video. For the last 18 months the top of the AI video leaderboard has been controlled by closed weights — Veo, Kling, Seedance, Runway — with the open ecosystem stuck a generation behind. Alibaba's Tongyi Lab shipped Wan 2.7 in April 2026 under Apache 2.0, packaged as a full suite: text-to-video, image-to-video, first/last frame control, and instruction-based video editing. The result genuinely competes with the closed flagships on resolution, motion fidelity, and audio — all while staying open-weight under Apache 2.0.

The architectural choice behind that is a 27-billion-parameter Mixture-of-Experts (MoE) model that activates only ~14 billion parameters per generation. The MoE design gives Wan 2.7 the capacity advantages of a much larger dense model — better world knowledge, better style coverage, better motion physics — while keeping inference latency closer to a 14B dense model. For high-volume production work (ad-tech generation, batch storyboarding, large content libraries), this is the model that scales.

The capability that sets Wan 2.7 apart functionally is instruction-based video editing. Pass an existing clip plus a text instruction — "make the jacket red," "change the background to a beach," "shift to golden hour" — and Wan applies the edit while preserving the original motion, character identity, and scene geometry. The closest competitor here is Veo's restyle pipeline, but Wan's instruction parser is more flexible. For agencies that have client footage they need to remix without re-shooting, this is the model that closes the loop.

Other notable capabilities: first/last frame control lets you lock the opening and closing frame and have Wan generate the bridging motion — useful for hitting specific narrative beats and for chaining clips that need to connect cleanly. Native synchronized audio — dialogue, ambient, and music — is generated with the picture in the same pass, lip-synced without a separate Foley step.

Where it's weaker: on top-end visual fidelity in single-shot work Veo 3.1 still has the edge at 1080p+ and the 4K upscale path. On multi-modal reference Elo Seedance 2.0 leads the public benchmarks. On explicit multi-shot storyboarding with hard cuts in one prompt, Kling V3 is more controllable. Wan's sweet spot is open-source friendliness, instruction edits, and batch production.

A reasonable mental model: Wan 2.7 is the default when you need open-weight provenance, high-volume batch production, or an editing-heavy workflow. For top fidelity, Veo 3.1. For reference-heavy shots, Seedance 2.0.