MiniMax H3 on an RTX 3060 12GB: what we actually measured

Quick answer

The card is not the bottleneck people expect. Our two runs finished at 2,579.8 and 2,581.8 seconds and both peaked at 11,649 MiB of 12,288 MiB VRAM — leaving 639 MiB of headroom. The constraint that actually decided whether the job could run at all was system memory, which peaked above 43,000 MiB. Full environment, hashes and per-run values are in the test card below.

43 min — Wall time for one 1344×768, 124-frame clip. Two runs: 2,579.8s and 2,581.8s.

94.8% — Peak VRAM, 11,649 of 12,288 MiB. Identical on both runs.

43,587 MiB — Peak system memory on run 1; 43,907 MiB on run 2.

The clip this card produced

Both runs used the same fixed seed and wrote a byte-identical file. The properties below were read directly from the MP4 container, not copied from the workflow settings.

Watch the 5.17-second output on YouTube

Machine-generated content. This poster frame and the linked clip are MiniMax H3 model output from our measured run SITE-3060-R2V-GPU0-B1, generated from two still reference photographs. The people, scene and vehicle plate are synthetic. Outputs of the model carry their own territorial restrictions under the MiniMax H3 Community License — read the license map before reusing them.

Verified output properties

• Resolution: 1344 × 768 (1.032 MP)
• Frames: 124 at 24 fps
• Duration: 5.166667 s — not an even 5 seconds
• Audio: no audio track
• File size: 1,317,884 bytes
• SHA-256: cc2a6f75…be6060f

Both runs produced this same hash. With a fixed seed and an unchanged workflow file, this configuration is deterministic on our machine.

This clip is entirely synthetic. Every frame was generated by MiniMax H3 from two still reference photographs. The people who appear in it do not exist, and the characters on the vehicle plate are model output, not a real registration.

Everything needed to check or repeat this

Version numbers and hashes are included because a MiniMax H3 timing is meaningless without them. Any of these can change whether the same workflow file runs at all.

One deliberate change from the official template: in the stock template the resolution inputs are driven by a ResolutionSelector node set to 0.4 megapixels, so the 1344×768 values visible in the node are not what actually runs. We disconnected that link and fixed the resolution directly, which is why our workload is 1.032 MP rather than the template default of 864×480.

Turbo LoRA note: the Turbo LoRA file named in several community workflows does not exist in the official Hugging Face repository at either the pinned revision or main. We did not substitute a different file, so our runs have no Turbo LoRA at all.

Run 1 and run 2 side by side

Our testing protocol requires any job over 30 minutes to be run twice, with both raw values published. We do not publish an average, because an average would hide how much these numbers actually move.

Observed range for wall time: 2,579.8–2,581.8 seconds.

Three findings that changed our own advice

Memory is set by the model, not the resolution

We ran a second workload on the same card and software that was roughly 40 times smaller — 512×288 at 22 frames instead of 1344×768 at 124. It finished in 55.6 seconds instead of 43 minutes, but peak VRAM only fell from 11,649 to 11,591 MiB and peak system memory only fell from 43,587 to 43,176 MiB. Both under one percent. Dropping resolution bought us time, not memory.

The card runs with 639 MiB to spare

At 11,649 MiB of 12,288 MiB, a 12GB 3060 is not comfortably inside its budget — it is sitting just under the ceiling. That headroom is small enough that another process holding a few hundred megabytes of VRAM, or a workflow addition that loads one more model, is a plausible cause of an out-of-memory error. Close other GPU applications before starting.

Two cards did not mean two jobs

We have two RTX 3060s in this machine and tried to establish whether both could work at once. They could not, and the reason was not VRAM. A single instance peaked above 43,000 MiB of system memory even at our smallest workload, so two would need roughly 86 GB. We ruled this out by measurement rather than by adding swap to force it through.

The first finding is the one with the widest consequences, and it has its own page: see why MiniMax H3 VRAM requirements barely move with resolution.

The run that froze the whole machine

Our first attempt at the full-size job did not produce a slow result. It produced no result, and took the host down with it. The sequence is worth publishing because the symptom looks exactly like an out-of-memory crash and is not one.

1. The job was submitted to the second of our two RTX 3060s and ran normally for about nine minutes.
2. That card reached 87–88 °C. Monitoring stopped mid-sample; the entire host became unresponsive and needed a manual power cycle.
3. Recovery was not clean either — the first restart failed because the GPU had dropped off the bus and needed physical intervention.
4. Afterwards we checked the obvious suspect and ruled it out: there was no out-of-memory entry in the system log, and the memory curve was flat rather than climbing toward exhaustion. The failure was a PCIe fatal completion timeout, not memory.
5. We then ran the identical small workload on each card in turn. Same job, same software: one card peaked at 60 °C, the other at 79 °C. A 19 °C gap that follows the physical slot, not the software.
6. Every number elsewhere on this page comes from the cooler card. The hot one is excluded from our benchmarks until it is physically addressed.

What this does and does not tell you

The 19 °C gap is measured. Our explanation for it is not: we think the two cards sit in adjacent slots with too little space, starving the second one of intake air. That is a hypothesis, and the measurement does not distinguish it from a fault in that card's own cooler, paste or power delivery.

It is testable in one run. Separate the cards physically, repeat the identical small workload on the previously hot card, and compare. Landing near 60 °C supports the spacing explanation; staying above 75 °C points at the card itself.

If your own dual-GPU machine locks up under sustained load, check temperatures per card with an identical job before assuming your workflow or your RAM is at fault.

How our run differs from the community reports

Reports of much shorter MiniMax H3 times on an RTX 3060 are real, and they are also not measuring what we measured. Four differences matter before any of the numbers below are read.

These are different workloads on different software stacks. Our figure and the community figures are not measurements of the same thing, and this page does not present them as a comparison, a validation or a contradiction of each other. We have not run the community configuration ourselves.

What our run does establish is a floor: this is what the official workflow costs on this card with no acceleration extras, which is the configuration most people actually start from.

These rows are deliberately not averaged and not ranked. They use different tasks, resolutions, model variants and software stacks, so a combined figure would be a fabricated benchmark.

A first run that tells you something

1. Check your system memory before your VRAM. On our machine the model needed over 43,000 MiB regardless of output size, so a 32GB system is the first thing to verify, not the GPU.
2. Start with the official template at its real default — 864×480 — rather than the resolution shown in the node, which the template overrides.
3. Keep audio off for the first run so a decode failure cannot be confused with a sampler failure.
4. Close other GPU applications. With 639 MiB of headroom, a second process holding VRAM is enough to change the outcome.
5. Record one successful baseline with a fixed seed before adding cache, LoRAs or longer output. Ours reproduced byte-for-byte, which makes any later change measurable.

Model folders and the official workflow JSON files are covered in the ComfyUI setup guide. If your run fails rather than runs slowly, start from the troubleshooting guide.

Two cards are still not one big card

Two RTX 3060s do not present 24GB to a workflow, and our attempt to at least run two jobs in parallel was blocked by system memory rather than VRAM. Multi-GPU only helps if a workflow explicitly distributes model work across devices, and the host still has to hold every instance's system memory at once.

RTX 3060 questions

Can an RTX 3060 12GB run MiniMax H3?

Yes. We completed the official reference-to-video workflow twice on our own RTX 3060 12GB. Each run produced a 1344x768, 124-frame clip in about 43 minutes, using 11,649 MiB of the card's 12,288 MiB.

How much system RAM does MiniMax H3 need on an RTX 3060?

More than the 32GB figure that circulates in community threads. Our peak system memory was 43,587 MiB on the first run and 43,907 MiB on the second, measured from the container cgroup. That is roughly 42.6 GiB, which exceeds the total capacity of a 32GB machine.

Does lowering the resolution reduce MiniMax H3 memory use?

Not meaningfully, in our measurements. We ran a workload roughly 40 times smaller on the same card and the same software: peak VRAM fell by 58 MiB and peak system memory fell by 411 MiB, both under one percent. Lowering resolution saved time, not memory.

Do two RTX 3060 cards give MiniMax H3 24GB of VRAM?

No. We tested whether our two cards could run two jobs at once and the answer was no, but the limit was system memory rather than VRAM. A single instance already peaked above 43,000 MiB even at our smallest workload, so two instances would need roughly 86 GB.