FusionCow releases Sulphur-3, a MiniMax H3 fine-tune for video with audio starting at 11.4 GB

For years, open video models have suffered from the same awkward workflow: one neural network generates the visuals, another handles the sound, and everything is then stitched together manually, often falling out of sync. That makes any release based on a model capable of producing video and audio in a single pass worth watching. FusionCow has delivered exactly that: the third version of its Sulphur fine-tune is built on MiniMax H3.

What exactly was released

Sulphur-2 was known as an option for generating video without built-in censorship. For the third version, the creator switched the base to MiniMax H3, which generates a clip together with its audio. The main weights are available in three formats: BF16 at 40.2 GB, INT8 at 20.9 GB, and W4A8 at 11.4 GB. This is the standard quantization ladder: the smaller the file, the lower the memory requirements and the greater the risk of quality loss.

There is also a separate Ref2VA variant with support for reference images. Put simply, you provide a photo or illustration of a character, and the model tries to preserve that character's appearance throughout the clip. This is one of the most persistent problems in video generation: without a reference, the face and clothing tend to drift from frame to frame. Ref2VA is available in BF16 and INT8, with the latter weighing 22.1 GB.

The creator has also published a 2.4 GB LoRA extracted from Sulphur-3. It is a convenient option if you do not want to download the full model and instead want to apply the changes to an H3 base you already have.

Setup and hardware

The repository includes an S3_ComfyUI.zip archive, while the reference-enabled variant runs through a dedicated Ref2VA node. One important caveat: the H3 encoder and video/audio VAE must be downloaded separately. This means the actual disk footprint will be larger than the combined size of the files listed above. If you plan to deploy the full setup on your own machine or a rented server, leave some headroom.

There is one user-reported performance figure. The INT8 Ref2VA variant, running on a GPU with 32 GB of VRAM, generated 124 frames at 672×384 in roughly 50 seconds using 20 steps and no corrector. The tester did not specify the exact GPU, so this is a reference point for one configuration rather than a benchmark. The resolution is modest: it is sufficient for previews and drafts, but final production will require upscaling.

Limitations and practical takeaways

The first limitation is access: files on Hugging Face are currently available by request with manual approval, and the creator has not announced when access will open to everyone. The second is that this is a third-party, uncensored fine-tune. If you are building a public-facing product, you will be responsible for moderating inputs and outputs, enforcing age restrictions, and managing the legal risks surrounding generated content, including the use of real people's likenesses in reference images.

What does this mean for production teams? First, treat the release as a signal: the combination of video and audio in a single model, together with open weights, has already reached community fine-tunes, which means more neutral variants are likely to appear soon. Second, if you are testing H3 in ComfyUI, consider W4A8 and INT8 as ways to fit the model into 12–24 GB of memory, but verify audio and facial quality on your own use cases. Third, for projects requiring a consistent protagonist, take a closer look at the Ref2VA approach: reference-based generation can be useful in both advertising and game-production pipelines.

The bottom line: this release is primarily interesting from a technical perspective, as an example of a recent open foundation model with audio and reference-image support. Until access becomes open and independent measurements are available, however, it should be treated as an experiment rather than a production-ready tool.