1 · Video memory
Ours peaked at 11,649 MiB of 12,288, leaving 639 MiB. Tight, but it completed twice without a single out-of-memory error. This is the budget that scales least with what you ask for: the model has to be resident regardless.
Quick answer
For the official REF2VA workflow, budget a 12GB card and — more importantly — well over 32GB of system RAM. We measured 11,649 MiB of VRAM and more than 43,000 MiB of system memory, and both figures stayed essentially flat across a 40× range of output sizes. Lower resolution makes the job faster, not smaller. Reports of MiniMax H3 running on 6GB and 8GB cards exist, but they use different model combinations and we have not verified them on our own hardware.
What we measured
The one tier we can speak to first-hand
Everything in this section is from our own machine. Every other tier on this page is somebody else's report, labelled as such.
11,649 MiBPeak VRAM of 12,288 available — 94.8%, identical on two independent runs.
43,587 MiBPeak system memory on the same run. The second run reached 43,907 MiB.
0.5%How much peak VRAM fell when we cut the workload by roughly 40×.
Hardware was an RTX 3060 12GB on Ubuntu 24.04.4 with the official REF2VA pruned INT8 model, no Turbo LoRA and no SageAttention, producing 1344×768 at 124 frames. The complete environment, version numbers, file hashes and both runs' raw values are on the RTX 3060 test card.
Three budgets, not one
VRAM is the number people ask about, not the one that stops them
A MiniMax H3 run has to fit three separate resources at once. Most guides discuss only the first. In our runs the first was the one with the most margin.
2 · System memory — the binding constraint
Ours peaked above 43,000 MiB on every run, including a workload that finished in 55 seconds. A machine advertised as 32GB cannot hold that at all. If you are choosing where to spend money for MiniMax H3 on a 12GB card, this is where it goes, and swap is not a real substitute.
3 · Disk
The REF2VA diffusion model alone is 20,970,379,616 bytes — roughly 19.5 GiB in one file — before the text encoder, VAE, cache, temporary files and outputs. We treat 45GB free as the floor for setting this up, and verified that before starting.
Why this ordering matters: the standard advice for a memory error is to lower the resolution. On our configuration that advice is close to inert for memory, because resolution is not what is consuming it. If you are hitting a hard limit at 32GB of system RAM, no resolution setting will get you under it.
The measurement
A 40× smaller job, and almost the same memory
We ran two workloads on the same card, same model files and same software, differing only in output size. Both used the same fixed seed and the same derived workflow file.
Same six numbers as the table below, plotted for scale. The system-memory panel includes a reference line at 32GB because that is the ceiling most community guides quote as sufficient.
| Workload | Relative size | Wall time | Peak VRAM | Peak system memory |
|---|---|---|---|---|
| 512×288 · 22 frames | 1× | 55.6 s | 11,591 MiB (94.3%) | 43,176 MiB |
| 1344×768 · 124 frames | about 40× | 2,581.8 s | 11,649 MiB (94.8%) | 43,587 MiB |
| Change | — | 46× longer | +58 MiB (0.5%) | +411 MiB (1.0%) |
What this shows: for this model and workflow, memory is dominated by resident model weights and working buffers that are allocated regardless of output dimensions. Time scales with the work; memory largely does not.
What it does not show: that this holds for every MiniMax H3 model variant, for audio-enabled runs, or past the point where a larger output would exceed the card. We measured two points on one configuration, not a curve.
Evidence by tier
VRAM requirements at each tier: what is actually known
Each row records one configuration that was reported to work, not a minimum for that tier. The confidence column is the point of the table: only one row is ours.
| VRAM | Reported hardware | Reported workload | Result | Evidence grade |
|---|---|---|---|---|
| 6GB | RTX 3050 6GB · 16GB RAM | Default 480p · 5 sec | Commenter reported about 17 min | Low — single comment, no environment. Source |
| 8GB | RTX 4060 Ti 8GB · 32GB RAM | FL2VA pruned INT8 · NVFP4 encoder · 640p · 5 sec | About 20 min cold; 12 min with EasyCache; about 95% VRAM reported | Medium — detailed post, but a different model combination than ours. Source |
| 12GB | RTX 3060 12GB · our machine | REF2VA pruned INT8 · 1344×768 · 124 frames · audio off | 11,649 MiB peak on two runs; 43 min each | Site-tested — full environment and hashes on the RTX 3060 page |
| 16GB | RTX 4070 Ti SUPER 16GB · 32GB RAM | INT8/NVFP4 I2V · 20 steps · 24fps · audio off | Five reported runs from 2:09 to 6:55; no OOM reported | High for a report — full environment and public JSON. Source |
| 20GB | RX 7900 XT 20GB · 32GB RAM | ROCm nightly · INT8/INT4 I2V · 1056×608 · 10 sec · 20 steps | About 58 min | Medium — detailed environment, no exact workflow file. Source |
| 24GB+ | Not normalized in this evidence set | More headroom, but still workflow dependent | No comparable measurement available | Unknown — we will not invent a figure |
Read the model column before the VRAM column. The 8GB report and our 12GB measurement are not in tension: they load different diffusion models and different text encoders. A smaller quantization can fit a smaller card, which is exactly why a single "MiniMax H3 needs X GB" number does not exist.
Note on the system RAM column: several of these reports pair a small card with 32GB of system memory. Our own measurements exceeded that figure on every run, so if you are planning a build from these rows, treat their RAM values as reported rather than as verified sufficient.
Memory model
Seven things that change the requirement more than resolution
Model precision
Full, INT8, NVFP4 and GGUF variants load different amounts and use different kernels. This is the single largest lever, and it is why the 8GB and 12GB rows above are not comparable. Smaller is not automatically faster or more compatible.
Which model, not just which size
REF2VA, FL2VA and the text-to-video variants are separate files with separate footprints. "MiniMax H3" is not one weight file.
Text encoder
An NVFP4 encoder and a full-precision one are a large difference in resident memory, and an out-of-memory error during encoding happens before sampling ever starts.
Frames, not seconds
Two clips both described as five seconds can carry different frame counts. Our 124 frames at 24fps is 5.166667 seconds, not five. Record frames and fps.
Audio
Audio adds models and a decode stage. We ran with audio off, so every figure on this page excludes it. Note that disabling the audio decode does not necessarily stop audio latents being generated during sampling.
Cache and offload
Cache can cut repeat time while consuming system RAM. CPU offload can fit VRAM by moving pressure onto system memory and storage — which, given the section above, is the budget you can least afford to spend.
Software versions
ComfyUI, PyTorch, CUDA and custom-node versions change whether a workflow runs at all, which makes any memory figure without version numbers difficult to act on.
Lowest-risk setup
Start small, then change one limit
- Confirm free system memory first, then free disk, then VRAM — in our runs that is the order in which they actually bind.
- Use the official template or a well-documented quantized workflow, and note which model variant it loads.
- Run once with audio off and a fixed seed to establish a baseline you can compare against.
- Close other GPU applications and record free VRAM before loading. Our headroom was 639 MiB.
- After one success, change either frames or resolution — not both — and expect the time to move far more than the memory.
FAQ
MiniMax H3 memory questions
How much VRAM does MiniMax H3 need?
On the one configuration we measured ourselves, 11,649 MiB of a 12,288 MiB card, or 94.8%. That figure is for the official REF2VA pruned INT8 model with audio off. Other model variants have different footprints, so there is no single number that covers every MiniMax H3 workflow.
Will a smaller resolution let MiniMax H3 fit in less VRAM?
Much less than most guides suggest. We ran the same model on the same card at a workload about 40 times smaller and peak VRAM moved from 11,649 MiB to 11,591 MiB. That is a saving of 58 MiB, or half a percent. Most of the memory is the model sitting in the card, and it is there whatever size clip you ask for.
Is system RAM or VRAM the real limit for MiniMax H3?
In our measurements, system memory was the tighter of the two. Peak system memory stayed above 43,000 MiB at every workload we ran, including the smallest one. A 32GB machine does not have that much in total, while the 12GB card completed the job with 639 MiB to spare.
Can MiniMax H3 run with 8GB of VRAM?
There is a detailed community report of a 4060 Ti 8GB completing a 640p five-second job, but it used a different model combination than the one we measured. We have not tested an 8GB card, so we grade that report as community evidence rather than a confirmed requirement.
Related guides