The home computer test is more revealing than the model announcement
There is a difference between downloading an open model and having a workflow that you can actually use. The first tells you that the files fit on a disk. The second tells you whether you can wait for a result, repeat a prompt and finish a project without fighting the hardware.
Our earlier MiniMax H3 test showed that an NVIDIA RTX 5060 Ti with 16GB of VRAM can generate a short video with synchronised audio in ComfyUI. A 1344 × 768 clip completed successfully, although the 5.17-second result took roughly 34 minutes in our configuration. That was enough to prove that local H3 video is possible on a consumer card. It did not prove that the same PC could behave like a small video production workstation.
What we tested next
For the follow-up, we used three photographs of a Ford Mustang Mach-E. The prompt asked MiniMax H3 to preserve the car while showing a stylus-driven technical explanation: a transparent body revealing an 88kWh NMC battery, a transparent front bonnet revealing a 203kW electric motor, and a final transparent rear section with rotating wheels and a 615km range label.
This is a more useful home test than a simple text-to-video scene. The workflow had to use visual references, maintain the identity of the same car, move between front and rear views and follow several ordered actions. We supplied the front and rear photographs as the first and last frames, while the side photograph guided the car’s proportions.
The result completed at 1344 × 768. The source MP4 was 10.125 seconds long at 24 frames per second. We removed the generated audio for this product-style test. It was not a perfect engineering animation, and labels produced by a generative model should never be treated as a source of technical truth. But the clip was good enough to demonstrate a practical use: a local user can turn a set of product photographs into a visual concept without uploading them to a cloud video service.
Full HD was the hard boundary
We then changed the target to 1920 × 1080. We tried 15, 12, 10 and 9 seconds. None of these native Full HD renders completed on the RTX 5060 Ti.
The failure was not caused by a lack of system memory, by a full queue or by another AI application consuming the card. We stopped the ComfyUI process, the transcription service and Ace-Step before repeating the experiment. We also restarted ComfyUI without its normal VRAM reservation. The failure still appeared in the SamplerCustomAdvanced node as a PyTorch out-of-memory error.
The most useful measurement came from the nine-second attempt. ComfyUI reported approximately 13.5GiB already allocated and another 1.02GiB requested during sampling. CUDA had only a few megabytes available at the critical moment. In other words, the problem was a temporary memory peak inside denoising, not the size of the final MP4 and not the amount of RAM installed in the computer.
This distinction matters. “MiniMax H3 runs on a 16GB GPU” is true for our tested workflow and resolution. It is not a universal promise that any 16GB card can generate any duration and any resolution. The model, quantisation, reference conditioning, frame count and sampler all affect the peak.
Why a second GPU and the CPU do not simply add memory
Our machine also has an RTX 3050 with 6GB of VRAM. It can run a separate workload, but ComfyUI does not automatically combine the two cards into a single 22GB pool for one H3 sampler. A second GPU is useful for another independent process; it does not make a 16GB card behave like a 22GB card.
CPU offloading is helpful for moving selected components out of the way, but it does not remove the largest temporary allocation made during the sampler’s denoising pass. Running the entire workflow on the CPU would technically change the device, but the resulting speed would make a long video impractical for most home users.
There is a more advanced idea called model parallelism, where layers are deliberately split across devices. That requires support in the model implementation, quantisation kernels and sampler. It is not a switch that can be enabled with CUDA_VISIBLE_DEVICES=0,1, and a six-gigabyte secondary card is not an attractive target for this kind of custom split.
Turbo reduces steps, not the laws of physics
We also tried the community MiniMax H3 Turbo node. It adds a Turbo LoRA and a dedicated sampler designed to reduce the usual roughly 20 sampling steps to four to eight. Its low_vram option merges the LoRA into the weights and is documented as a way to lower peak memory on quantised models.
That is a useful optimisation. It can reduce waiting time and may make a difficult resolution possible. It is not, however, a larger graphics card. With the Turbo LoRA, four steps and its low-VRAM mode enabled, our nine-second native Full HD attempt still ended with an out-of-memory error.
For a home user, Turbo is therefore worth trying at 1344 × 768. It is not a guarantee that 16GB can produce native 1920 × 1080. The model’s output quality can also change: fewer steps and a merged LoRA may trade some sharpness for a lower memory peak.
What can a European home user realistically do?
MiniMax H3 is attractive in Europe because local generation can keep source photographs and prompts on the user’s own computer. That can be useful for a small business preparing an unreleased product, a school working with student material or a creator who does not want every reference image sent to a third-party server. Local processing does not remove every GDPR responsibility, but it changes the data path: there is no automatic upload to a hosted video provider in the workflow we tested.
The practical choices are straightforward:
- Use 1344 × 768 for generation. This is the mode our RTX 5060 Ti handled with both text and image references.
- Upscale after generation. A 1344 × 768 clip can be converted to a 1920 × 1080 delivery file. That produces a Full HD file, but it does not create native Full HD detail inside H3.
- Use Turbo at the lower resolution. Four to eight steps may make prompt iteration less painful, with a possible quality trade-off.
- Buy more VRAM only for a clear reason. A 24GB or 32GB GPU offers more headroom, but changing from NVIDIA CUDA to AMD ROCm also means validating a different software stack.
AMD’s documentation now lists the Radeon AI PRO R9700 with 32GB and ROCm support, while AMD’s MiniMax H3 article demonstrates an H3 deployment through ROCm and SGLang on Instinct GPUs. Those are promising signals, not a tested drop-in replacement for our ComfyUI setup. Anyone considering an R9700 should budget time for a separate ROCm environment and compatibility testing.
The honest verdict
A 16GB NVIDIA card makes sense for MiniMax H3 if your goal is learning, storyboarding, short product concepts or local experiments. It can generate real video, use image references and keep the source material on the machine. That is already useful.
It becomes the wrong tool when “Full HD” means native 1920 × 1080 output from the H3 sampler, especially with reference frames and a technically demanding prompt. We tried longer and shorter durations, freed the other AI services, removed the VRAM reservation and added Turbo. The same boundary remained.
For a home user in Europe, the sensible workflow is not to discard a 16GB card. Generate at the tested lower resolution, upscale for delivery and treat every iteration as a local laboratory run. If the project requires native Full HD, fast retries and predictable production, the deciding specification is not the model’s marketing label. It is the available VRAM headroom.
Sources
MiniMax H3 source repository, ComfyUI discussion of MiniMax H3 memory behaviour, ROCm compatibility matrix and AMD’s MiniMax H3 ROCm deployment article. The Ford Mach-E results and VRAM measurements are our own test results on an RTX 5060 Ti with 16GB of VRAM.
Can I use MiniMax H3 locally without sending product photos to a cloud service?
Yes, with the open local H3 workflow used in our test. The images and prompt can remain on the computer, although the owner still has to handle local storage, access control and any other GDPR obligations that apply to the project.
Does a 16GB graphics card produce a Full HD file?
It can produce a Full HD delivery file after upscaling, but our native 1920 × 1080 H3 renders failed with out-of-memory errors. Native generation worked at 1344 × 768 in our configuration.
Is a 32GB AMD GPU automatically compatible with this workflow?
Not automatically. The extra memory should provide much more headroom, but AMD requires a ROCm-based PyTorch and ComfyUI environment. The exact H3 nodes, quantisation kernels and sampler must be tested separately from the NVIDIA setup.
