OpenMOSS drops MOVA to generate synced video and audio

MOVA is an open-source foundation model from the OpenMOSS team that generates video and synchronized audio together in a single pass.

Instead of the old clunky pipeline where you generate silent footage first and then try to paste an AI voiceover or lip-sync model on top, MOVA handles both modalities natively inside the same model, cutting down on cumulative errors.

What it does

MOVA handles precise multilingual lip-sync alongside environment-aware sound effects.

In official demos, it manages everything from a formal indoor speech to a multi-person conversation between a man and a child walking through a park, matching vocal cadences, room acoustics, and mouth movements cleanly.

On audio synchronization benchmarks, it reportedly outperforms LTX2. Unlike closed-source competitors like Sora 2, Veo 3, or Kling, OpenMOSS went fully open-source by releasing weights, inference code, training pipelines, and LoRA fine-tuning scripts.

Under the hood and hardware requirements

The model relies on an asymmetric dual-tower architecture that connects pre-trained video and audio representations via bidirectional cross-attention.

It is built as a Mixture-of-Experts (MoE) network with 32 billion total parameters, but only 18 billion parameters are active during inference to keep computing manageable.

OpenMOSS released both 360p and 720p versions. The catch is hardware: the 720p model weighs around 77 GB, which is way too heavy for typical consumer GPUs.

To make local testing feasible, the repo includes CPU offloading options. For generating an 8-second 360p clip, layerwise group offloading drops VRAM usage to 12 GB on an RTX 4090, though it demands 76.7 GB of host RAM and runs at roughly 42 seconds per step. On an H100 with component-wise offload, it takes 48 GB of VRAM and runs at 9 seconds per step.

How to run it

All code and documentation are live on GitHub under the OpenMOSS/MOVA repository, with weights hosted on Hugging Face.

You can clone the repo, set up a Python 3.13 environment, and test sample inference scripts for single-person and multi-person prompts. Third-party ComfyUI nodes and API access are also available if you prefer not to manage raw checkpoints.

Why it matters

Most open-source video generators have been essentially mute, forcing creators to rely on fragmented tools or closed commercial platforms for sound and speech. MOVA changes that by giving the community full access to native audiovisual weights and training code.

That said, it is not a casual download. Running 720p requires serious enterprise hardware, and even squeezing 360p onto an RTX 4090 requires massive system RAM and plenty of patience. We think it is a major technical milestone for open video, but practical daily use will depend on how quickly the community can quantize and optimize the pipeline.