TongFlow Now Integrates MiniMax-H3 — Accelerated with Sol-Attn

TongFlow now integrates MiniMax-H3. The official tongflow-modal-minimax-h3 plugin brings the best open-weight omni-modal video model to the canvas — text, images, videos, and audio all go in as references, and a video with native stereo sound comes out.

What you get

MiniMax-H3 is a 33B parameter model that unifies six video generation modes into one node family:

• Text / image → video — the classic case, with optional first-frame or first+last-frame control.
• Omni-reference video (refs-gen-video) — feed up to 9 images, 3 video clips, and 3 audio clips at once, and the model weaves them into a single new clip.
• Voice-driven video — pair a portrait with an audio track and get a video that matches.

On the canvas, this shows up as a natural multi-select action: pick a few reference images, videos, or a prompt, and TongFlow surfaces “Reference video” right in the toolbar. No manual node wiring needed.

Fast out of the box

Running a 33B model isn’t free — but it doesn’t have to be slow either. We built the plugin on top of Sol-Attn, NVIDIA’s training-free sparse attention method published in late July, using kijai’s open-source Triton implementation. Instead of computing full attention over every token, it dynamically routes to the blocks that matter and approximates the rest — no fine-tuning, no LoRA, no quality loss we could detect.

The result: a 5-second clip that used to take about 5 minutes now finishes in under 3 minutes, with the same visual quality. All of it is already tuned and shipped in the plugin — you don’t touch a single parameter.

Thank you, open-source community

None of this exists in a vacuum. MiniMax open-sourced the H3 weights days after launching it as an API product. NVIDIA Labs published Sol-Attn as a paper with working code. kijai — one of the most prolific ComfyUI plugin authors — turned that paper into a usable node within days. We just wired it all together and made it one click away. If you build in the open, thank you; this is what makes tools like TongFlow possible.

Start for free

The plugin runs on Modal, which gives every new account up to $30/month of free GPU compute on B200-class hardware. That’s enough headroom to generate real clips without owning a GPU. You’ll need one credential in TongFlow Settings: MODAL_TOKEN_ID and MODAL_TOKEN_SECRET.

Try it

1. Install the plugin — open the plugin manager in TongFlow and install tongflow-modal-minimax-h3. No restart needed.
2. Add your Modal credentials in Settings.
3. Drop a prompt, a few images, or a video clip on the canvas, pick a video mode, and run. The platform auto-deploys to Modal the first time you use it.

This is the whole point of TongFlow: state-of-the-art open-source models, wired up and ready to drag onto a canvas, with the acceleration work already done for you.

New to TongFlow? Start with the introduction, grab the desktop app from Releases, or jump straight into the hosted studio at app.tongflow.com. If this is useful, a star on GitHub helps a lot.

Expand your imagination, stretch your ideas — give it a try.