Huge thanks & support in TongFlow now
Huge thanks to the MiniMax team for open-sourcing.
Thanks also to Comfy-Org for the day-0 optimized weights (pruned int8 ConvRot), which bring the full system from 498 GB down to ~63 GB and make single-GPU self-hosting practical.
We've just shipped MiniMax-H3 support in TongFlow — an open-source, node-based multi-modal AIGC studio.
TongFlow doesn't proxy or resell inference. You bring your own Modal token (BYOK);
Modal currently includes up to $30/month of free GPU compute, and we measured ~4 min 17 s per 5s clip on a B200 (≈ $0.45/clip) — so a few dozen clips a month fit inside the free credits. Both task checkpoints (FL2VA + Ref2VA) plus the NVFP4 text encoder stay resident in the B200's 192 GB, so switching modes costs nothing.
Everything is open source — the studio, the plugin, and the Python SDK. If you try it and something breaks, issues and PRs are very welcome.
Once again — congratulations on the release, and thank you for pushing open omni-modal generation forward. 🎉
🏠 TongFlow (open-source studio): https://github.com/tong-io/tongflow
🔌 MiniMax-H3 plugin: https://github.com/tong-io/tongflow-modal-minimax-h3