Ovi

Ovi answers the Veo 3 question for local hardware: one open model producing video and its soundtrack together. Character AI's release runs an 11B twin-backbone system in which a video branch and a 5B audio branch exchange information during generation, so speech lands on lip movements and effects land on events without a separate foley pass. Clips run 5 or 10 seconds at 24fps in aspect ratios up to a 960x960 pixel budget since the v1.1 update of November 2025, driven by text alone, text plus image, or image-to-video, with spoken lines marked inline in the prompt. Everything is Apache-2.0: weights in safetensors including an FP8 variant, single and multi-GPU inference code, and a Gradio UI. Hardware needs come to 32GB VRAM with CPU offload, dropping to 24GB using FP8 quantization, while 80GB cards run unconstrained; community ComfyUI support arrived through the WanVideoWrapper nodes. For talking characters and sound-on shots generated without cloud APIs, it is currently the leading open option, and the repo stays active with regular updates.