Running MiniMax H3 Locally: Quants, VRAM and Real Render Times
H3 has been runnable on consumer hardware since the weights went out in August 2026, and the setup guides all agree on the steps. Where they disagree — or stay quiet — is on the numbers: which quant fits which card, and how long you will actually wait. This page collects the reported figures in one place.
✦ See the unlimited plan
Pick the quant before you pick anything else
The released checkpoints come in two families: fl2va for text- and image-to-video, and ref2va for reference-conditioned generation. Both are published as GGUF quantizations, and the file you pick decides everything downstream.
The original BF16 checkpoints are about 123.6 GB, which is why nobody runs them at home. Community quants bring a working setup down to roughly 21 GB.
One warning worth repeating: the Q8_CR and U16G builds depend on the ComfyUI-GGUF custom nodes and will not load in other runtimes.
| Quant | fl2va | ref2va | Notes |
|---|---|---|---|
| Q4_0 | 11.4 GB | 11.4 GB | Standard 4-bit. Smallest fit, lowest quality. |
| U16G | 15.0 GB | 15.0 GB | Mixed INT8 + Q4_0. Faster than Q4_0 on 16 GB+ cards. |
| Q8_CR | 20.2 GB | 20.2 GB | Q8_0 quality with INT8 ConvRot weights. |
| Q8_0 | 20.2 GB | 21.4 GB | Standard 8-bit. Highest quality of the four. |
What each card can do, and how long it takes
These are reported figures from the launch period rather than a controlled benchmark, and they move with offload strategy, storage speed and how much system RAM is free. Read them as an order of magnitude, not a promise.
The pattern underneath them is consistent, though: below 16 GB you are trading minutes for fit, and above 24 GB the canvas opens up before the clock does.
| VRAM | Example cards | Workable canvas | Time per 5s clip |
|---|---|---|---|
| 8 GB | RTX 3070 · 4060 · 3060 Ti | 640×360 – 832×480 | 9–10 min (RTX 3070) |
| 12 GB | RTX 3060 · 4070 · 2080 Ti | 0.4 MP default | under 9 min (RTX 3060, 20 steps) |
| 16 GB | RTX 4070 Ti Super · 4080 · 5070 Ti | 0.4–0.5 MP; 768p with Turbo LoRA | ~3 min (4090 Laptop, 960×540) |
| 24 GB | RTX 3090 · 4090 | Full 1344×768 canvas | Low single-digit minutes |
| 32 GB | RTX 5090 | Full canvas, up to 15 s | ~10 min (20-step GGUF Q5) |
Buy RAM before you buy VRAM
This is the one piece of advice that comes up in every low-VRAM writeup, and it is counter-intuitive enough to be worth stating plainly: when the model does not fit, ComfyUI streams the rest from system memory. A 12 GB card with 32 GB of RAM finishes; the same card with 16 GB thrashes or freezes the machine.
Plan on 32 GB of system RAM as the floor for a 12 GB card, and 64 GB if you are on 16 GB and want it to feel smooth. A fast NVMe matters for the same reason — the tensors that leave the GPU have to come back from somewhere.
Two optimizations beat all the others
Turbo LoRA is the big one. It drops the step count to 8 for text- and image-to-video and 4 for reference-to-video, and it cuts render time by more than any change of quantization will.
Sage Attention is the second, via the `Patch Sage Attention KJ` node from ComfyUI-KJNodes. It lowers peak VRAM in the self-attention layers, which is what turns a card that OOMs into one that finishes.
Everything else is a smaller lever. Get a plain workflow running first, then add one of these at a time — an optimization added before you have a baseline is indistinguishable from a broken install.
| Technique | What it does | When to reach for it |
|---|---|---|
| Turbo LoRA | 8 steps (T2V/I2V), 4 (R2V) | Always. Biggest single win. |
| Sage Attention | Lower peak VRAM in attention | When you OOM during sampling. |
| Dynamic offloading | Streams parts to system RAM | When the checkpoint will not fit at all. |
| KJNodes low-VRAM attention | Lower peak VRAM, identical output | Same case, no quality cost. |
| More system RAM | Makes offloading survivable | Before any GPU upgrade under 16 GB. |
The settings to start from
Begin at the small canvas and move one variable at a time. The defaults below are the official ones, and changing them before you have a clean run is how an afternoon disappears.
One thing that looks like a bug and is not: the frame count lands on H3’s `17k+5` alignment, so a 5-second request becomes about 124 frames — 5.17 seconds at 24 fps.
| Setting | Value |
|---|---|
| ComfyUI | 0.30.0 or newer |
| Test canvas | 864×480 (0.4 MP) |
| Native target | 1344×768 (0.98 MP) |
| Steps | 20 (official default) |
| Sampler | res_multistep |
| Scheduler | simple |
| Output | 24 fps video + 32 kHz stereo audio, MP4 |
When it breaks, it breaks in the same few ways
Most first-run failures are one of these, and none of them mean the model is wrong for your machine.
| Symptom | Usual cause | Fix |
|---|---|---|
| Nodes or templates missing | ComfyUI older than 0.30.0 | Update, then restart completely. |
| Model not in the loader | Wrong folder or filename | Check `diffusion_models/`, `text_encoders/`, `vae/`; restart. |
| R2V fails, T2V works | fl2va loaded instead of ref2va | Select the ref2va checkpoint. |
| CUDA out of memory | Canvas or duration beyond the plan | Back to 0.4 MP, 5 s, batch 1. One variable at a time. |
| Machine freezes | Offload pressure beyond system RAM | Close other apps, free disk, retry the baseline. |
| R2V far slower than T2V | Too many or too large references | Set `ref_image_size: match`, cut the reference count. |
| Video saves silent | Audio VAE not wired to CreateVideo | Feed both video and audio VAE outputs into it. |
Render small, upscale the keepers
This is not a low-VRAM compromise — it is how H3 is built. The base model renders at its native size and a separate pass takes it up; MiniMax split those two jobs because pixels are far cheaper to add afterwards than to generate.
The same split is the right habit on any hardware. Most takes are discards — the fourth attempt, the one where the hand is wrong — and putting every take through an upscale spends the expensive half of the pipeline on footage you will delete. Generate at the small canvas, watch them, upscale the ones you keep.
If the hardware is the problem
Everything above assumes a card worth several hundred dollars sitting idle while it renders, for minutes per clip. If that is not your situation, the practical route is a hosted service that does not charge per clip — because the render-and-discard loop is exactly what makes per-clip pricing hurt.
HelloGen is the best option we know of for that: unlimited H3 generation on the Pro plan, 0 credits per clip, with at least 2 priority renders a day and standard speed beyond that. Same render-small-then-upscale habit, without the machine. The arithmetic behind that split is in why no free AI video generator exists.
Sources
Every figure on this page came from one of these, and none of it is our own measurement — we run H3 hosted, not on the hardware described above.
· ComfyUI — MiniMax H3 video generation for the node names, the 0.30.0 requirement and the official workflow settings. · MiniMax-H3-GGUF on Hugging Face for the quantization file sizes and the ComfyUI-GGUF dependency. · awesome-minimax-h3 for the collected VRAM figures and runtimes. · Community write-ups for the per-card render times and the Turbo LoRA / Sage Attention effects — individual reports from the launch period, which is why this page calls them reports rather than benchmarks.
Frequently asked
What VRAM do I need to run MiniMax H3?
8 GB works with a Q2/Q3 GGUF and 32 GB of system RAM, at around 9–10 minutes for a 5-second clip. 12 GB is comfortable at 0.4 MP. 24 GB runs the full 1344×768 canvas. Below 16 GB, system RAM matters as much as the card.
Which MiniMax H3 GGUF quant should I download?
U16G at 15.0 GB is the sweet spot on 16 GB and above — mixed INT8 and Q4_0 weights, reported faster than Q4_0. Q8_0 (20.2 GB fl2va, 21.4 GB ref2va) is the quality pick; Q4_0 at 11.4 GB is the smallest fit. Q8_CR and U16G need the ComfyUI-GGUF nodes.
How long does a 5-second H3 clip take locally?
Reported times run from about 3 minutes on a 16 GB laptop 4090 at 960×540 with Sage Attention, to 9–10 minutes on an 8 GB RTX 3070. Turbo LoRA cuts this more than any quantization change.
What are the official workflow settings?
20 steps, res_multistep sampler, simple scheduler, 864×480 to test and 1344×768 as the native target, on ComfyUI 0.30.0 or newer. Output is 24 fps video with 32 kHz stereo audio.
Why does my 5-second clip come out as 124 frames?
H3 aligns frame counts to 17k+5, so a 5-second request lands at about 124 frames — 5.17 seconds at 24 fps. Expected behaviour, not a bug.
Can I use H3 without a local GPU?
Yes — through a hosted service. HelloGen includes unlimited H3 generation on the Pro plan at 0 credits per clip, with at least 2 priority renders a day.