HyperFlow H3 Flash: GGUF Setup, VRAM and Video Costs

For PCs with 64GB system RAM or less, Video Rebirth recommends the GGUF route through ComfyUI, Unsloth Studio or patched stable-diffusion.cpp. The published desktop tests cover 24GB and 32GB GPUs, including a fast-disk configuration on a PC with 16GB of system RAM. A small weight file alone does not establish that a 16GB GPU will work.

The Python INT8 package targets data-center GPUs or RTX machines with much more host RAM. With CPU offload, the documented requirements are roughly 91GiB of system RAM for the bf16 text encoder, or 69GiB with fp8. Without offload, it needs about 90GiB VRAM. Choose the runtime before choosing a precision file.

Both routes generate 5.2–15-second clips at 24fps, defaulting to 1344×768, with 32kHz stereo audio. FL2VA covers text, first-frame and first-plus-last-frame tasks. Ref2VA uses a different transformer: GGUF supports image references, while Python INT8 also handles video and audio references, within the documented per-type limits and twelve-reference total.

For FL2VA, Q4_K_M is a 9.1GB transformer file; Ref2VA’s is 9.2GB. The recommended desktop setup also downloads a 15.7GB Qwen3VL NVFP4 text encoder and the required video and audio VAEs. Q8_0 transformers are 14.9GB and 15.2GB respectively. Select all the components required by your workflow, rather than downloading the transformer and stopping there.

On 16GB or 32GB host-RAM PCs, the official guide starts ComfyUI with --fast-disk, streaming unresident weights through the operating-system file cache. Use fast local storage and allow space for the companion components, downloaded variants and output clips. The publisher warns that a 15-second run can exhaust a 16GB host without that setting.

For the publisher’s Q4_K_M GGUF setup at 1344×768 and six steps, peak VRAM is 23GiB for a 5.2-second clip and 22GiB for a 15-second clip on 24GB cards. The RTX 5090 measurements are 28GiB and 27GiB respectively. These are complete-runtime peaks, separate from transformer download sizes.

INT8 has a separate first-run download: about 90GB with the bf16 text encoder or 68GB with fp8. The package downloads the selected compressed transformer, text encoder and VAEs, without fetching the full 62GB base H3 transformer. First-use Triton compilation adds about a minute in the documented setup.

Begin with a working ComfyUI installation and a compatible GPU backend. Install ComfyUI-GGUF and Video Rebirth’s HyperFlow Flash nodes in custom_nodes, then follow their dependency and component-download instructions. The nodes provide dedicated templates; a generic MiniMax H3 graph can miss the compressed model’s conditioning and sampling behavior.

Download the matching FL2VA Q4_K_M file, the documented 15.7GB text encoder and both required VAEs to the locations in the integration guide. Restart ComfyUI, then open Workflow → Browse Templates → ComfyUI-HyperFlow-Flash and choose hyperflow_t2v for a first text-prompt clip. Select the matching model files and keep the supplied six-step schedule.

Try a five-second scene with one subject and simple sound, such as a paper boat drifting across a quiet pond with water sounds. Inspect the saved video and audio, not just the preview frames. For endpoint-frame control, switch to hyperflow_i2v or hyperflow_flf2v. hyperflow_ref2v requires the separate Ref2VA weights.

The publisher’s six-step GGUF table uses 1344×768 and includes text encoding, transformer sampling and video/audio decoding. With 32GB host RAM, the RTX 5090 32GB takes 118 seconds for a 5.2-second clip and 618 seconds for 15 seconds. The RTX 4090 24GB takes 151 and 815 seconds respectively. GPU memory, host RAM, precision and clip duration all belong in the comparison.

The advertised 92-second RTX 5090 result uses the Python INT8 route and its larger host-memory requirement. The 8.7-second result uses four B200 GPUs to produce a 5.2-second clip with the Python INT8 package. NVIDIA SOL-H3’s partner runtime is a separate acceleration route. These configurations have different memory needs from the GGUF desktop setup.

In an independent November 7 hands-on report, KOBO Lab uses Windows 11, an RTX 5090 32GB, 128GB RAM, ComfyUI 0.44 and HyperFlow node commit 20a22ab. Its Q4_K_M, six-step, roughly five-second 768p warm run takes about 107 seconds; the first load-and-generate takes 3 minutes 40 seconds. These timings come from a different environment than the publisher’s table.

BennyDaBall’s unofficial ComfyUI INT8 loader is a third route alongside publisher GGUF and Python INT8. Its author reports a five-seed mean of 46.1 seconds from queue to saved MP4 on an RTX 5090, for 1344×768, 124 frames, six steps and audio; peak VRAM is 24.2GB. The same test records 66 seconds for official GGUF nodes. This uses converted INT8 checkpoints, native convrot kernels and the official HyperFlow node pack. The report does not specify host RAM.

Phantom’s November 8 report runs GGUF on a 16GB RTX 5060 Ti in a dual-card PC with 64GB RAM; the second GPU uses only 114MiB. With CPU offload and disk-mapped weights, six-step reference clips at 672×928 and 124 frames take 316.3 seconds for Q4_K_M and 307.1 for Q8_0. Reported process RAM is 31.3/37.1GiB, with 12,236–12,264MiB reported peak GPU use. This narrower-resolution case expands the tested options, while the publisher’s own desktop table remains a 24/32GB-GPU comparison.

The Python package requires Python 3.10 or newer, PyTorch 2.6 or newer with CUDA and Triton, diffusers 0.40 or newer, and an Ampere-or-newer NVIDIA GPU. Multi-GPU execution requires Linux or WSL2. The fp8 extra enables the lower host-memory text-encoder path; check the complete requirements before running the command below.

For text-to-video, supply a prompt and output filename. The CLI also accepts --image and --last-image for endpoint frames, or ordered --reference inputs with Ref2VA. A requested frame count is rounded up to the supported 17n+5 grid; use the documented range when setting longer clips. Keep duration, reference mix, step count and offload mode with every timing result.

The same repository now includes fl2va_bf16 (27.4GB) and ref2va_bf16 (27.9GB), the compressed few-step transformers before INT8 quantization. These folders support re-quantization or conversion to another format. They remain HyperFlow H3 Flash derivatives, with the MiniMax H3 weight-license conditions.

The Python package and its AutoModel implementation can load these folders. ComfyUI’s nodes instead read GGUF or the single-file ComfyUI safetensors. The publisher recommends INT8 or GGUF for generation; its published speed and memory tables measure those formats. Add the text encoder, VAEs, buffers and cache when sizing a BF16 run.

When converting, retain the two-time conditioning and the listed float32 input, output and time-embedding components. The repository provides the weights without training code or conversion scripts.

For a hypothetical compatible GPU rental at $2 per hour, the publisher’s 815-second RTX 4090 GGUF run would allocate about $0.45 for a 15-second output. Three attempts would cost about $1.36 before loading, idle time, storage and finishing. The hourly rate is an example, not an available rental quote; the elapsed time belongs to the publisher’s specific configuration.

If the machine is already yours, calculate electricity from measured whole-system power and elapsed time, then account separately for hardware ownership. For an assumed 0.5kW draw over 815 seconds, energy is about 0.113kWh. At an assumed $0.20/kWh, that is roughly $0.023 per attempt before hardware costs.

Compare against a hosted route using the same duration, resolution, audio and references. Pika H3 Flex offers a separate checkpoint and hosted finishing; the older HyperFlow package is an eight-step LoRA on the full-size H3 transformer. Match the actual workflow rather than choosing only by the number of sampling steps.

The compressed weights remain derivatives governed by the MiniMax H3 Community License. Its standard grant excludes the United States, European Union, United Kingdom and South Korea without prior authorization from MiniMax, and includes an acceptable-use policy. Establish the applicable license before installing the weights for a service or local project.

Commercial product or service interfaces must prominently display MiniMax H3. If your commercial products and services generate more than $20 million in yearly revenue, obtain separate prior written authorization from MiniMax. Redistribution requires passing on the agreement and, except for hosted services, its NOTICE file. Publicly disseminated generated content must be clearly labeled as machine-generated under the acceptable-use policy.

Video Rebirth’s Python code is Apache-2.0, its ComfyUI nodes use GPL-3.0 and the stable-diffusion.cpp integration is MIT. These software licenses apply to different code components; they do not replace the model-weight agreement. HyperFlow H3 Flash is published by Video Rebirth and is not an official MiniMax product.