Fewer sampling steps do not automatically make a large video model practical: memory footprint and the cost of each pass matter just as much. The key idea here is to combine a shortened 4–8-step trajectory with structural compression, bringing a roughly 31B-parameter base transformer down to 13.7B or 13.9B parameters for its two task partitions.
Key Capabilities
- It produces video and synchronized stereo audio together, with text-to-video, keyframe-guided generation, and reference-guided generation. This makes it useful for workflows where sound and visuals need to be generated as one output.
- The compressed transformer is available in rotated W8A8 INT8 and several GGUF quantizations. The Q4_K_M weights are about 9.1–9.2 GB, offering a lower-memory option for consumer GPUs; INT8 targets different runtimes and hardware configurations.
- On a 5.2-second, 768p clip at six steps, the reported end-to-end time is 92 seconds on one RTX 5090 and 8.7 seconds on four B200 GPUs. Those figures are not real-time; real-time performance on four Blackwell GPUs is associated with NVIDIA’s separate SOL-H3 runtime.
- The
ref2vapath can accept images, videos, and audio in the Python implementation, while GGUF runtimes support image references only. This distinction matters when choosing a runtime for mixed-media input.
Who It Suits—and the Trade-Off
Great fit if you want to experiment with joint video-and-audio generation and can match a runtime to your GPU and memory budget. Look elsewhere if you need turnkey real-time performance on a single consumer GPU, broad reference-media support across every runtime, or a permissively licensed model: the weights follow the MiniMax H3 Community License Agreement, which imposes territorial and usage restrictions.
