Wan

Wan did for video what Llama did for language: Alibaba released a frontier-adjacent video model family as open weights — including a 5B variant that runs on a consumer GPU — and the open-source ecosystem (ComfyUI workflows, LoRAs, fine-tunes) materialized around it within weeks. Output reaches native 720p, with 14B models pushing 1080p, and quality focuses on the hard parts: physical motion, human articulation and prompt adherence.

Distinctive controls matter to creators: first/last-frame conditioning lets you steer from one pose to another — a storyboard tool more than a slot machine. Costs drop to electricity for local users; the tradeoffs are the usual open-model ones — slower iteration per clip than hosted rivals and the need for your own VRAM discipline. For studios and creators who want video generation they own, Wan is the current default.