HyperFlow H3 Flash: 14B Transformer, Minutes on RTX 4090

What was actually cut

This is not a quantization of the original model with a new name. According to the model card, the fl2va transformer drops 3 of its 50 blocks and factors 102 linear layers; the ref2va transformer drops 1 block and factors 114 layers. Another 180 linear layers are quantized to rotated 8-bit integer form, and a two-time conditioning scheme, where every token carries both the current noise level and the step endpoint, is what allows the model to sample in 4 to 8 steps without classifier-free guidance. The Python route downloads HyperFlow's approximately 16 GB compressed INT8 transformer rather than the roughly 62 GB upstream MiniMax-H3 transformer. The text encoder, video VAE, audio VAE and other required components are downloaded separately. The parameter reduction comes from structural pruning and low-rank factorization; quantization further reduces storage and runtime requirements.

The tasks survive the cut on paper. Text-to-video, first-keyframe and first-and-last-keyframe generation, and reference-to-video with up to 12 references are all listed as supported. One limit sits inside that list: in the GGUF runtimes most consumer users will actually install, reference-to-video accepts images only. Video and audio references need the Python package, with its much larger memory demands.

The speed table, read by hardware

The fastest figure in HyperFlow's published benchmark table is 8.7 seconds to generate a 5.2-second clip at six sampling steps, using four NVIDIA B200 GPUs. That is approximately 1.7 seconds of generation time per second of finished video, not real-time output. The project separately attributes real-time performance on four Blackwell GPUs to its partner inference stack, NVIDIA SOL-H3. The model card does not provide an equivalent end-to-end timing table for that separate runtime.

The consumer rows tell a different story, and to its credit the project publishes them in full, end to end, including text encoding and decoding. On an RTX 4090, Video Rebirth reports 157 seconds for a 5.2-second clip using GGUF Q4_K_M on a system with at least 64 GB of RAM. Its disk-streaming GGUF configurations take 151–153 seconds with 16–32 GB of RAM. The fastest listed 4090 result, 130 seconds, uses the Python INT8 runtime with streaming offload, which requires approximately 69 GiB of system RAM with the FP8 text encoder or 91 GiB with BF16. These are different deployment configurations, not interchangeable speed settings. A full 15-second clip takes 830 seconds in GGUF, close to fourteen minutes. An RTX 3090, the card many local-video hobbyists actually own, needs 266 to 315 seconds for the short clip depending on runtime, and roughly 28 to 31 minutes for the long one. The GGUF short-clip benchmark reports approximately 23 GiB of peak VRAM on 24 GB-class cards, leaving little headroom. The Python INT8 streaming-offload configuration uses less VRAM for that clip but requires substantially more system RAM.

None of this makes the release a failure. A model that needed datacenter hardware now completes, locally, on a gaming PC, with its soundtrack attached. It does mean the honest unit for consumer hardware is minutes per clip, and any summary that leads with 8.7 seconds is describing a machine the reader almost certainly does not have. All of these figures are also the project's own. No independent benchmark of Flash appeared in the sources checked for this piece.

The 9.1 GB file is not the install

The second number travelling with the release is 9.1 GB, and it is accurate about exactly one file: the compressed transformer in Q4_K_M form. The setup instructions on the same card add the rest. ComfyUI users also download a 15.7 GB text encoder, plus the video and audio VAEs. The Python INT8 route downloads about 90 GB on first run, or about 68 GB with a smaller text encoder, and its offload modes call for 69 to 91 GB of system RAM. The GGUF route has the lower stated system-RAM requirement: Video Rebirth documents configurations with 16 GB of RAM using ComfyUI's --fast-disk option. That does not make 9.1 GB the complete download or total memory footprint. The Q4_K_M transformer is accompanied by a 15.7 GB text encoder, video and audio VAEs, and runtime dependencies. Disk-streamed execution also depends on storage performance, which the published benchmarks do not fully characterize.

The card is also plain about the bottom of the range: 16 GB graphics cards have not been measured yet. Readers with one should treat support as untested, not implied by the headline file size.

What the card does not measure

For a release built on removing more than half a transformer's parameters, the missing document is a quality comparison. The card benchmarks time and memory in detail and says nothing measurable about what the pruning and quantization did to the picture. The card says the derivative inherits MiniMax-H3's prompt adherence, supported languages, biases and failure modes, but publishes no matched quality evaluation against the full-size model. There are no reported side-by-side scores for visual fidelity, temporal consistency, prompt following or audio synchronization under equivalent generation settings. That leaves the quality cost of pruning and few-step sampling unknown. It does not establish that the compressed model performs worse, or that it preserves the original model's quality.

The license deserves the same close reading as the benchmarks. HyperFlow's transformer weights are distributed under the MiniMax H3 Community License, while the main GitHub inference code is separately licensed under Apache-2.0. The weights are not licensed for the European Union, United Kingdom, South Korea or United States without prior authorization from MiniMax. The license also requires prior written authorization for products or services exceeding $20 million in annual revenue, prominent “MiniMax H3” identification in products built with the weights, and disclosure when generated content is publicly distributed. Redistribution and hosted services carry additional obligations, and the license restricts using outputs to train or improve other AI models. An Apache-2.0 code badge does not override the model-weight license. Two bundled integrations carry their own licenses: the ComfyUI custom node is GPL-3.0 and the stable-diffusion.cpp patch is MIT.

Sources

[1] Video Rebirth, HyperFlow H3 Flash GitHub repository (created 6 November 2026)Read source
[2] Video Rebirth, HyperFlow H3 Flash model card, Hugging Face (read in full 8 November 2026)Read source
[3] Atlas360, "Video Rebirth cuts MiniMax-H3 model below 14B parameters" (7 November 2026)Read source

THE ALFW VERDICT

HyperFlow H3 Flash is a meaningful compression effort, not merely a smaller quantized copy of MiniMax-H3. Video Rebirth documents structural pruning, low-rank factorization and few-step sampling, and its 9.1 GB Q4_K_M transformer makes local experimentation more accessible. The speed claims need their hardware attached. The 8.7-second result uses four B200 GPUs, while the project's RTX 4090 figures range from 130 to 157 seconds for a 5.2-second clip, depending on runtime and system memory. The fastest consumer result also requires substantial RAM for offloading. Those are useful engineering trade-offs, not consumer real-time generation. The unanswered question is quality. Video Rebirth publishes detailed speed and memory tables but no matched comparison showing what the compression and reduced sampling steps cost against full MiniMax-H3. Try it for the smaller footprint and local video-and-audio generation. Do not assume it preserves the base model's output quality, or repeat the 8.7-second headline without naming the four B200s.