MiniMax H3's default files in ComfyUI add up to about 42.5 GB, more than any consumer graphics card holds. (This guide's default set is the template's video model and text encoder with the full-size fp16 video decoder; the template's own 8-bit decoder is 2.4 GB smaller.) It still runs on 24 GB cards, and on 16 GB cards with some care, because ComfyUI does not need everything on the card at once.
This guide shows what takes the memory, how ComfyUI shares it out, what decides render time, and what to choose on 16, 24 and 32 GB cards. For which files to download, see H3 model file names decoded; for what each file does, MiniMax H3 in ComfyUI: what every part does.
A few words first
- VRAM (video memory): the memory on the graphics card. The card can only compute on what is there.
- RAM (system memory): the computer's main memory. ComfyUI keeps loaded files there.
- Offloading: keeping part of a model in RAM and moving it to the card when it is needed.
- Step: one round of cleaning noise out of the clip. A clip takes about 20, and each runs the whole video model.
- GB: gigabyte. File sizes are counted in thousands of millions of bytes; memory is counted in binary gigabytes, which are about 7% larger. So a 24 GB card holds about 25.8 GB of files.
What takes the memory
H3 needs four files. Stacked together, they are larger than any of these cards, and only 64 GB of RAM holds them all with room to spare.
- The video model (21.0 GB) is the part that runs at every step, so it matters most. On a 16 GB card even this file alone does not fit, and part of it is streamed from RAM at each step. On a 24 GB card the file fits with little room left for the clip being worked on, so ComfyUI still streams part of it. A 32 GB card holds it with room to spare.
- The text encoder (15.7 GB) runs once per prompt, before the steps start. It takes its turn on the card, then waits in RAM.
- The two decoders (5.8 GB together) run once at the end.
- Not shown: the clip being worked on also takes memory on the card, more for longer and larger clips. No maker publishes a figure for it, which is why the chart shows files only.
How offloading works
When the video model does not fit in VRAM, ComfyUI keeps the rest in RAM and streams those weights to the card during every step.
That sounds slow, and on paper it is a lot of traffic. In practice it costs little. ComfyUI's H3 implementer, Kijai, has explained why: the transfer runs in the background while the card computes, and each H3 step involves so much computing that the transfer is almost entirely hidden. His conclusion is that a smaller file only helps when the model is larger than your VRAM and RAM combined.
That gives the one rule that matters most:
All the files a workflow loads should fit in system RAM at once, with room to spare.
If they do not, users report that parts are read from disk again for each job, and a render that should take minutes takes many times longer.
How much RAM you need
Add up the files you load, then add your operating system, ComfyUI itself and anything else you have open.
| Set | Files | Comfortable with |
|---|---|---|
Default (8-bit video, 4-bit encoder, fp16 decoder) | about 42.5 GB | 64 GB of RAM |
| 8-bit video, 8-bit encoder | about 54 GB | 64 GB, tight; more is better |
| 16-bit video (pruned bf16), 4-bit encoder | about 62 GB | 96 to 128 GB |
| 16-bit video, 16-bit encoder | about 98 GB | 128 GB or more |
6-bit video (w6a8), 4-bit encoder, 8-bit decoder | about 35 GB | 48 to 64 GB |
The last column is arithmetic plus headroom, not a measurement. With 32 GB of RAM the default set does not fit, and more RAM is the most useful upgrade for H3, ahead of a bigger card.
The text encoder costs RAM, not time
The text encoder's full-precision file is 51.5 GB, but it runs once per new prompt or start picture, never once per step. Render a second seed of the same prompt and ComfyUI reuses its answer. A larger encoder costs RAM, and a few seconds whenever the prompt changes.
What drives render time
Five things decide how long a clip takes, roughly in this order:
- Steps. Each step runs the whole video model, so time grows almost in step with the count. That is why 4- and 8-step speed-up LoRAs save so much.
- Length. The model's attention relates every part of the clip to every other part, so a clip twice as long takes more than twice as long.
- Size. The same holds for resolution. ComfyUI's page says it plainly: lower megapixel values run faster.
- The attention method. By ComfyUI's account, Sage Attention roughly doubles speed, at the risk of morphing late in a clip (see the settings guide).
- Memory. If the files do not fit in RAM, everything else is swamped by reading from disk.
A faster card finishes each step sooner, of course. But two machines with the same card can differ many times over if one is short of RAM. The first render after starting ComfyUI is also slower, because it includes loading the files.
32 GB cards (RTX 5090)
A 32 GB card holds the default 21 GB video model with room for long clips at full size, so 15-second clips at 20 steps are practical.
The 40 GB 16-bit video model also becomes a reasonable experiment, if you have 96 GB of RAM or more. MiniMax releases its weights in 16-bit, but ComfyUI recommends the 8-bit ConvRot file, and the 16-bit file is slower.
24 GB cards (RTX 4090, RTX 3090)
The default set runs at full quality.
Part of the video model lives in RAM and is streamed each step, which is what ComfyUI is designed to do. With 64 GB of RAM the default set fits comfortably. The 8-bit text encoder brings the files to about 54 GB, which is tight on a 64 GB machine. The RTX 3090 has the 8-bit integer units the default files use.
16 GB cards (RTX 4060 Ti, 5060 Ti, 4080)
H3 runs, with part of the video model streamed from RAM at each step.
The default 21 GB video model does not fit in 16 GB, but ComfyUI streams the rest, so it still works. What changes is the practical size and length. Community guides written for 16 GB cards try ideas at about 480p with fewer steps, then make only the chosen shots at full size.
- Have 64 GB of RAM. It matters more than anything else here.
- Keep the default 8-bit video model if your RAM allows; the 6-bit
w6a8file (16 GB) is the fallback. - Try ideas small: 0.4 to 0.5 MP or a speed-up LoRA while you explore. Remember that a different size is a different clip.
- Prefer shorter clips. Two 5-second clips render faster than one 10-second clip.
- Close everything else. Browsers with many tabs and other programs that use the graphics card take memory you need.
Under 16 GB
People run H3 on 12 GB and even 8 GB cards with 4-bit and GGUF files, low sizes and heavy speed-ups. It is possible, but the result is a compromise of all of those at once, and no maker publishes settings for it.
A checklist for a slow machine
- Is PyTorch built for CUDA 13.0 (
cu130)? ComfyUI recommends the defaultint8_convrotfiles only with it. - Do all loaded files fit in RAM? Add them up.
- Are models staying loaded between runs? A launch option that clears the model cache after each job makes every render load them again.
- Is another program using the card? Check before blaming H3.
- Is the slow render the first one after a restart? That one includes loading.
When the machine is ready, Your first H3 video: a checklist goes from an empty screen to a finished clip.
Sources: Kijai's comments in the discussions of Kijai/MiniMax-H3-experimental; the file listing of Comfy-Org/MiniMax-H3; ComfyUI's H3 tutorial; MiniMax's H3 model card; the 16 GB deployment guide at neng320/minimax-h3-local-deployment.