h3 studio
video + audioNative Metal video and audio generation from MiniMax-H3, no PyTorch in the loop.
helmstudio installs, runs and switches between studios — each one a repository that wraps an open model — and keeps what every studio makes in one gallery, with where it came from.
There is no release yet, so helmstudio runs from a source build.
A manifest names the repository, the commit and whether it has submodules, and install clones all of it at that commit, then runs the build steps the manifest lists.
A studio declares the Python it needs, and helmstudio gives it an environment of its own, made with uv, so two studios never share one.
Weights come from Hugging Face a file at a time, and a download resumes where it stopped. Studios that name the same repository at the same revision share one copy, and a folder you already have can be linked instead.
A port in a manifest is a preference: when something else has it, the process is given a free one. And one studio holds a model in memory at a time, so launching a second says what it would stop, and asks.
What each one needs, read from its manifest.
Native Metal video and audio generation from MiniMax-H3, no PyTorch in the loop.
MLX video and audio generation against a user-selected LTX checkpoint.
Native Metal image generation across the FLUX.2 Klein and Z-Image checkpoints.
PyTorch speech and voice generation on mps, with AuK and AuK-Flash checkpoints.
Most repositories worth running will never ship a manifest, so writing one for a repository you did not write is a first-class act, not a workaround.