FREE WEBINAR X NetApp
Quick Answer
MiniMax H3 can run on a 24 GB GPU, but that proves feasibility, not production readiness. Practical performance depends on quantization, offloading, resolution, and inference optimization. L40S provides 48 GB VRAM, while RTX PRO 6000 Blackwell provides 96 GB. The key decision is not fit or speed, but which setup delivers acceptable video quality at the lowest cost per finished minute.
MiniMax H3-Base weights can be deployed locally using supported runtimes such as SGLang, vLLM, Diffusers or ComfyUI. The official full-quality 2K validation workflow is hybrid: H3-Base can run locally, while MiniMax’s H3-Context-IR preprocessing and H3-Regenerate-2K stages are currently accessed through MiniMax APIs. A fully local H3-Base workflow produces the released 768p path. Developers can implement their own context-processing logic using MiniMax’s prompting guidance. However, getting the model to load is only the first hurdle. The questions that matter in practice are much more useful:
- How much VRAM does H3 really need?
- How long does a usable video take to render?
- What quality do you give up with quantization?
- How much compute does each finished minute cost?
Those questions matter because MiniMax H3 is an omni-modal generation system rather than a conventional image model. MiniMax officially supports 4 to 15-second outputs at 24 FPS with 32 kHz stereo audio. H3-Base generates at a 768-pixel short edge, while the complete 2K workflow uses a separate H3-Regenerate-2K stage.
Can MiniMax H3 Actually Run on Your GPU?
Yes. But, ‘can run’ and ‘runs well’ are both very different standards.
A practical MiniMax H3 deployment has to pass several stages.
- First, the model has to load.
- Next, it needs to finish inference without crashing or exhausting memory.
- Then it needs to render at a speed that is practical.
- Finally, the resulting cost has to make sense for repeated video generation.
A documented community test shows the difference clearly. On a single RTX 3090 with 24 GB VRAM and roughly 31 GB system RAM, an optimized ComfyUI setup generated 362 frames at 832×480, producing 15.083 seconds of video at 24 FPS. Peak VRAM was approximately 19.8 GB, but the render required 23 minutes 17 seconds.
Consequently, 24 GB should be treated as proof of possibility, not proof of production readiness.
The official Comfy-Org model repository also shows how much model packaging and quantization can change the memory picture. Its H3 diffusion-model files include a 66.3 GB BF16 checkpoint, a 34 GB INT8 ConvRot checkpoint, and a 21 GB pruned INT8 ConvRot checkpoint for both FL2VA and Ref2VA variants.
That does not mean a 21 GB checkpoint needs only 21 GB of VRAM during inference. Activations, text encoding, VAEs and runtime overhead still consume memory. Nevertheless, the file sizes make one point clear: checkpoint choice can materially change what hardware is practical.
How can system RAM become the real bottleneck?
VRAM is only part of the memory story.
In the RTX 3090 experiment, ComfyUI’s default pinned-memory behavior consumed 29,866 MB of roughly 31,997 MB host RAM and triggered the system OOM handler. Disabling pinned memory reduced measured host-RAM use to 7,508 MB, allowing the same 362-frame workload to complete.
However, this should not be treated as a universal RAM optimization. Disabling pinned memory changes CPU-to-GPU transfer behavior and can affect throughput, so benchmarking on the actual offload path is important before production use.
For H3, size GPU VRAM and system RAM together, especially when offloading is involved.
For higher-memory deployments, NVIDIA specifies the L40S with 48 GB GDDR6 ECC and 864 GB/s memory bandwidth. The RTX PRO 6000 Blackwell Server Edition provides 96 GB GDDR7 ECC and 1,597 GB/s bandwidth.
More VRAM provides valuable headroom. Nevertheless, twice the VRAM does not imply twice the H3 render speed.
How Should You Benchmark MiniMax H3?
A GPU comparison only becomes useful when both GPUs run the same MiniMax H3 workload. That sounds obvious.
First, identify which workflow is being tested. MiniMax separates H3-Base-FL2VA, which covers text-to-video and first/last-frame generation, from H3-Base-Ref2VA, which uses reference images, videos and/or audio.
The 2K path also needs separate treatment. MiniMax states that H3-Base produces 768p output, after which H3-Regenerate-2K feeds that result and the original context through another generation stage. Importantly, H3-Regenerate-2K is not currently open-sourced and is exposed through MiniMax’s API.
Therefore, a fair L40S vs RTX PRO 6000 comparison should use two tests. First, run an apples-to-apples benchmark with the same checkpoint, precision and runtime settings wherever both GPUs support them natively, while keeping the generation mode, prompt, seed, resolution, frame count, diffusion steps, attention backend, caching and software versions identical. Second, run a best-achievable-production benchmark using the optimal supported precision, kernels and acceleration path for each GPU. Report both results separately.
For production planning, record wall-clock generation time, peak VRAM, peak host RAM, finished video seconds per GPU-hour, retry rate and actual billed compute.
Software optimization alone can materially change those measurements. In a reproducible RTX PRO 6000 Blackwell A/B test, an 864×480, 192-frame workload fell from 182.324 seconds with Native + SageAttention to 104.178 seconds with TE-Speed + SageAttention, a 42.9% reduction. At 1344×768, generation fell from 514.150 seconds to 288.094 seconds, a 44.0% reduction.
The GPU did not change. The inference path did. Consequently, the apples-to-apples test isolates hardware differences, while the best-achievable-production test shows the performance users can realistically deploy. Reporting both prevents a software advantage from being mistaken for a hardware advantage.
How Fast Does MiniMax H3 Actually Render?
There is no trustworthy universal answer such as ‘MiniMax H3 takes five minutes.’ Instead, render time depends on resolution, frame count, quantization, attention implementation, offloading, caching, and inference acceleration.
Inline Research measured an L40S producing a 10-second clip at 960×544 in about 7.2 minutes. The same workload peaked at 38.9 GB VRAM and 46.7 GB system RAM, leading the team to recommend 64 GB host RAM.
At that exact throughput, 7.2 minutes per 10 seconds of output translates to roughly 83 seconds of finished video per GPU-hour. This is a derived throughput calculation, not a separate benchmark.
A separate reproducible RTX PRO 6000 Blackwell test reported:
| Workload | Measure Render Time |
|---|---|
| 864×480, 192 frames, Native + SageAttention | 182.324 sec |
| 864×480, 192 frames, TE-Speed + SageAttention | 104.178 sec |
| Reported time reduction | 42.9% |
| 1344×768, 192 frames, Native + SageAttention | 514.150 sec |
| 1344×768, 192 frames, TE-Speed + SageAttention | 288.094 sec |
| Reported time reduction | 44.0% |
The repository includes benchmark metadata, workflows, logs, and representative outputs, making the results more auditable than an unsupported performance claim.
These results should not be directly compared with the L40S number because the workloads differ. The stronger takeaway is that H3 performance depends on the GPU and inference stack together.
How Much Does Quantization Save, and What Does It Cost in Quality?
Quantization is probably the most useful lever for making MiniMax H3 easier to run. However, the tradeoff is not as simple as ‘lower precision means lower quality.’
Unsloth MiniMax H3 quantization project reports that its 40.23 GB pruned BF16 source falls to 20.25 GB with its quantised variants. Meanwhile, measured end-to-end render peak memory drops from 57.11 GB to 36.97 GB.
That is a meaningful reduction. More importantly, it shows why checkpoint size alone is not enough when planning GPU memory. Runtime overhead and other model components still matter.
The same project compared INT8 and FP8 against BF16 using matched prompts, seeds, shapes, and 20-step Ref2VA runs.
| Precision | Prompt 1 SSIM | Prompt 2 SSIM |
|---|---|---|
| INT8 | 0.92 | 0.82 |
| FP8 | 0.88 | 0.78 |
| BF16 repeat-reference ceiling | 1.00 | 1.00 |
The tests used the same prompt, seed, and shape at 20 steps. More importantly, the repository explicitly describes the SSIM figures as a divergence measure rather than a quality score.
Therefore, an SSIM of 0.92 does not mean INT8 retains ‘92% of BF16 quality.’ Production evaluation should still inspect faces, fine textures, temporal stability, motion coherence, prompt adherence, reference consistency, and audio-video synchronization.
Which quantization tier should you start with?
| Goals | Sensible starting point |
|---|---|
| Reference-quality validation | BF16 |
| Practical memory-quality balance | INT8 |
| Hardware-specific low-precision optimization | FP8 |
| Maximum Blackwell memory efficiency | NVFP4, after visual validation |
NVFP4 should not be treated as an equal option across L40S and RTX PRO 6000. The measured H3 NVFP4 implementation requires NVIDIA Blackwell for its native path. On Ada, Hopper and older architectures, the repository recommends INT8 instead.
That matters because RTX PRO 6000 is Blackwell, while L40S belongs to the Ada generation.
A separate RTX PRO 6000 Blackwell community benchmark measured a pruned INT8 model at 21.0 GB, 19,995 MB staged VRAM and 2.17 seconds per iteration, compared with 12.5 GB, 11,944 MB and 1.90 seconds per iteration for NVFP4. The measured difference was about 8.0 GB less staged VRAM and 12.4% lower sampling time.
A practical starting point is INT8 when output quality remains important, followed by lower-bit formats when memory pressure or throughput becomes the larger constraint.
What Does One Finished Minute ofMiniMax H3 Video Cost?
GPU hourly price tells only half the story. A cheaper GPU can still generate more expensive video if it takes substantially longer.
A more useful metric is:
₹ per accepted finished video minute = total billed instance compute ÷ accepted output seconds × 60
AceCloud offers a 1× L40S with 16 vCPUs and 64 GB RAM at ₹142.12/hour or ₹83,000/month. Applying that live hourly rate to Inline Research’s independently measured 7.2-minute runtime for 10 seconds of H3 output gives:
₹142.12 ÷ 60 × 7.2 ≈ ₹17.05 per 10-second output
Therefore:
₹17.05 × 6 ≈ ₹102.33 per finished video minute
However, ₹102.33 is a derived planning figure, not an AceCloud-measured H3 benchmark. It combines a reproducible third-party L40S runtime with AceCloud’s current public rate.
For RTX PRO 6000, AceCloud provides a 1× 96 GB configuration with 16 vCPUs and 128 GB RAM at ₹121,599/month.
Consequently, converting that monthly amount into a supposedly measured ₹/minute figure would require an assumed utilization rate. That would be an estimate.
The comparison worth publishing is an identical H3 workload on both GPUs with actual wall-clock time and billed compute recorded.
Want the actual ₹ per finished minute for your workload? Talk to an expert about benchmarking the same H3 prompt, resolution and precision on the GPU configuration you intend to deploy.
Should You Run MiniMax H3 Locally or on a Cloud GPU?
Self-hosting does not necessarily mean buying a GPU. The better question is whether hardware ownership or control of the inference environment is the priority.
Local or on-prem hardware is attractive when a suitable GPU is already available, utilization is consistently high, the team can maintain the inference stack, or strict local data control matters.
Cloud GPUs become more attractive when demand is bursty, 48 GB or 96 GB VRAM is required without purchasing hardware, or multiple GPU tiers need to be tested before committing capital. Those capacities come directly from NVIDIA’s L40S and RTX PRO 6000 Server Edition specifications.
Cloud economics should also include storage and data-transfer requirements when reference videos, generated clips or intermediate assets are large.
For H3, this distinction matters because a 24 GB GPU can demonstrate feasibility, whereas a higher-memory setup may provide a much more practical production envelope.
Therefore, the cheapest option depends on utilization, throughput, reliability and quality requirements, not GPU purchase price or hourly rate alone.
Which MiniMax H3 Setup Should You Choose?
MiniMax H3 is not simply a VRAM-sizing problem. It is a cost-performance problem.
A 24 GB GPU proves the model is accessible, but the documented RTX 3090 example still required 23 minutes 17 seconds to generate 15.083 seconds of video. L40S offers a more practical 48 GB memory envelope, while RTX PRO 6000 Blackwell Server Edition doubles that capacity to 96 GB. Meanwhile, quantization and software acceleration can materially change both memory demand and throughput.
Therefore, VRAM is the gatekeeper, but cost per acceptable finished output is the real decision metric.
A sensible production decision should balance:
quality + throughput + reliability + actual compute cost
rather than ranking GPUs by VRAM, theoretical performance or hourly price alone.
AceCloud lets teams access higher-memory GPU infrastructure without committing to hardware upfront. However, the right configuration should still be validated against the actual H3 mode, precision, output resolution and expected utilization.
Book a free consultation to validate the GPU, quantization tier and expected cost per accepted finished minute for your MiniMax H3 workload.
Frequently Asked Questions
Yes. A documented RTX 3090 workflow generated 15.083 seconds at 832×480 and 24 FPS, peaking at about 19.8 GB VRAM. However, the render took 23 minutes 17 seconds, so fitting H3 and running it economically are different questions.
There is no single requirement because offloading behavior changes the result. In one 31 GB RTX 3090 system, pinned-memory behavior caused an OOM failure, while disabling pinned memory cut measured host-RAM use from 29,866 MB to 7,508 MB and allowed the same workload to complete.
Not universally. In one matched Ref2VA test, INT8 produced SSIM values of 0.92 and 0.82, versus 0.88 and 0.78 for FP8. The benchmark explicitly treats those numbers as divergence measurements rather than subjective quality scores.
Not through the currently released full system. MiniMax says H3-Base produces 768p output, while H3-Regenerate-2K handles the 2K regeneration stage. That module is not yet open-sourced and is currently used through MiniMax’s API.
Combining the independently measured 7.2-minute L40S runtime for 10 seconds of output with AceCloud’s current ₹142.12/hour price gives approximately ₹102.33 per finished minute for that workload. It is a derived planning figure rather than a first-party AceCloud H3 benchmark.
