Video Rebirth Squeezes MiniMax H3 Flash Into a Single 24 GB GPU
Video Rebirth compressed MiniMax H3 from 32B to 14B with INT8 quantization and 6 inference steps, enabling real-time 768p video generation on 4x B200 GPUs and local inference on consumer RTX cards.
- HyperFlow H3 Flash compresses MiniMax H3 from 32B to ~14B parameters with INT8 quantization
- Generates 768p video + stereo audio in 8.7 seconds on 4x B200, 92 seconds on RTX 5090
- Uses 6-step sampling without classifier-free guidance via two-time (t, r) conditioning
- Runs locally in ComfyUI on 24 GB consumer GPUs with Q4_K_M GGUF at 9.1 GB
- NVIDIA SOL-H3 enables real-time generation on 4x Blackwell GPUs from day zero
- License excludes EU, UK, South Korea, and US without prior MiniMax authorization
Video Rebirth has released HyperFlow H3 Flash, the third model in its HyperFlow series. The release compresses MiniMax H3 to roughly 14 billion parameters, applies INT8 quantization, and generates synchronized video and stereo audio in four, six, or eight sampling steps.
Video Rebirth reports an 8.7-second end-to-end generation time for a 5.2-second clip on four B200 GPUs. Quantized GGUF weights also support local inference on a single 24 GB RTX card, although the reported RTX 5090 generation time rises to 92 seconds.
A 32B-class model, cut to 14B
MiniMax H3 is a 33-billion-parameter multimodal model that accepts text, images, video, and audio. It produces video and native 32 kHz stereo audio in one pipeline, which removes the synchronization work required when separate models generate the picture and soundtrack.
The H3 transformer contains roughly 31 billion parameters and occupies about 62 GB in BF16 precision. Its full-attention architecture also increases memory and compute requirements as the video context grows, placing the original weights beyond a typical single-GPU workstation.
HyperFlow H3 Flash reduces the first-frame-to-video-and-audio variant, called fl2va, to 13.7 billion parameters. The reference-to-video-and-audio variant, ref2va, contains 13.9 billion. The smallest Q4_K_M GGUF file is 9.1 GB, though total runtime memory also includes text encoders, VAEs, activations, and framework overhead.
Six steps, measured end to end
Video Rebirth measured the full pipeline at 1344×768 resolution, including text encoding, six transformer steps, and video and audio VAE decoding. The published results cover a 124-frame clip lasting about 5.2 seconds at 24 fps.
| Hardware | Latency | Latency per output second |
|---|---|---|
| 4× B200 | 8.7 seconds | 1.7× |
| 1× H200 | 36 seconds | 6.9× |
| RTX 5090 | 92 seconds | 17.7× |
| RTX 4090, 24 GB | 157 seconds | 30.2× |
| RTX 3090 | 315 seconds | 60.6× |
The four-GPU B200 path uses NVIDIA SOL-H3, an optimized inference implementation. Its 8.7-second result remains slower than playback for the measured clip, since real-time generation would require a result at or below 5.2 seconds.
Generated clips can run from 5.2 to 15 seconds. Hardware, quantization, input mode, frame count, and software backend can all change latency, so the published figures provide configuration-specific results rather than cross-platform guarantees.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
