When MiniMax H3 became available as an open-weight model, my first question was probably the same as yours:
Can I actually run this locally on my GPU?
I expected to find a simple minimum VRAM number. Instead, I found reports that seemed to contradict each other.
One person could run H3 with 8GB of VRAM. Another ran out of memory with 12GB. Some RTX 3060 users reported five-second generations in under ten minutes, while others needed much longer.
After reading through the details, the reason became clear: the GPU model alone does not tell us enough.
Two MiniMax H3 runs are only comparable when they use roughly the same:
A five-second clip at 0.4 megapixels is not equivalent to a ten-second clip at a higher resolution. An INT8 diffusion model with a quantized text encoder is also very different from the full weights.
This is why statements such as “H3 needs 12GB” or “H3 works on 6GB” should be treated as starting points, not universal requirements.
I organized the community reports I could find by GPU, RAM, model variant, workflow and generation settings.
These are community-reported results, not controlled benchmarks performed by me.
There are individual reports of H3 completing short jobs on 6GB cards.
However, the evidence is currently limited. Some results appear only in comments and omit important details such as exact model files, dependency versions, peak system RAM and workflow configuration.
I would not describe 6GB as generally supported based on the available evidence.
One of the more detailed reports used:
The author reported approximately 20 minutes for a cold run and around 12 minutes with EasyCache, while using roughly 95% of the available VRAM.
That proves one carefully configured 8GB setup can work. It does not prove that every 8GB GPU or workflow will work.
The strongest collection of consumer-GPU reports currently comes from the RTX 3060 12GB.
Several users with 32GB of system RAM reported short generations around:
Other 12GB workflows took considerably longer, particularly when system RAM, duration, resolution or workflow type changed.
My cautious interpretation is that 12GB is currently the clearest practical starting point for an optimized local H3 workflow, but it is not a guarantee.
A detailed RTX 4070 Ti SUPER report used an INT8/NVFP4 image-to-video setup with 20 steps and audio disabled.
Five reported runs took between roughly 2 and 7 minutes, depending on the input size and output configuration.
This is stronger evidence because the author provided the software environment and a public workflow. Even so, it remains one machine and one software stack.
An RX 7900 XT user reported completing a 1056×608, ten-second image-to-video run in approximately 58 minutes using a ROCm nightly environment.
The run completed, but it also shows why additional VRAM does not automatically mean faster inference. GPU architecture, backend support and software maturity matter too.
If I were setting up MiniMax H3 on a 12GB consumer GPU, I would begin with:
The goal of the first run is not maximum quality. It is to establish a clean baseline that successfully loads the encoder, diffusion model and VAE.
After that, change one variable at a time.
If you immediately add a custom workflow, LoRA, cache system, audio generation and higher resolution, it becomes much harder to identify which component caused an error.
Resolution matters, but an out-of-memory error can happen before sampling begins.
Possible failure points include:
If the workflow fails before the sampler starts, reducing the sampler resolution may not address the actual problem.
Watch both VRAM and system RAM, and record the stage at which memory usage spikes.
A reproducible H3 report should contain something like this:
GPU and VRAM:
System RAM:
Operating system:
GPU driver:
ComfyUI version or commit:
Python / PyTorch / CUDA or ROCm:
Workflow type:
Exact diffusion model:
Exact text encoder:
VAE files:
Resolution:
Frames and FPS:
Steps:
Sampler and scheduler:
Audio enabled:
Cache or acceleration:
Cold or warm run:
Generation time:
Peak VRAM:
Peak system RAM:
Without these details, “it works on my GPU” is useful encouragement, but weak technical evidence.
The guide I built
I turned my notes into a free MiniMax H3 VRAM guide with the configurations separated by evidence quality:
https://www.minimaxh3tutorial.com/vram
Full disclosure: this is my own independent site. There is no signup, paywall or affiliate link. Performance figures link back to their original community reports, and results that I have not reproduced are labeled as community-reported.
I am still treating this as a living evidence table, not a finished benchmark database.
One final license note
MiniMax H3 is available as open weights, but that does not automatically mean unrestricted open-source use.
The model has a Community License with territory and usage conditions. Check the current official license before down or using the weights for a project or client.
Corrections are welcome
If you have run MiniMax H3 locally, I would be interested in your exact configuration—especially AMD, Linux and lower-VRAM results.
Please include the model variant, workflow, resolution, frames, steps, RAM and timing. I would rather correct the table than preserve a neat but inaccurate conclusion.