Video Rebirth released HyperFlow H3 Flash, a compressed 13.7–13.9B parameter version of the 31B MiniMax-H3 transformer that jointly generates 5.2–15 s video at 1344×768/24 fps with synchronized 32 kHz stereo audio. The model achieves this through structural pruning (dropping 3 of 50 blocks for fl2va, 1 for ref2va), low-rank factorization of 102–114 linear layers, and rotated W8A8 INT8 quantization (convrot) applied to 180 layers behind a block-diagonal Hadamard rotation. A two-time (t, r) conditioning scheme injects the current noise level t and step endpoint r into every token, enabling 4–8 step sampling without classifier-free guidance. Inference runs at 6 steps default (4 for speed, 8 for quality). Benchmarks show 8.7 s for a 5.2 s clip on 4× B200, 36 s on 1× H200, and 92 s on 1× RTX 5090 (INT8, --offload stream). NVIDIA SOL-H3 enables real-time inference on 4× Blackwell GPUs. The INT8 transformer (~16 GB) targets data-center GPUs (H100/H200/B200/B300) and 24 GB+ RTX cards with ≥69 GiB system RAM (fp8 text encoder) or ≥91 GiB (bf16); GGUF quantizations (Q3_K_M–Q8_0, 7.2–15.2 GB) target consumer GPUs with ≥16 GB system RAM. Multi-GPU scaling uses Ulysses context parallelism across 1–8 GPUs with FSDP2 text-encoder sharding and parallel VAE tile decoding. First-run downloads ~90 GB (or ~68 GB with fp8 text encoder); the 62 GB base transformer is never downloaded. Reproducibility is bitwise-identical given matching seed, architecture, GPU count, and attention backend. Code is Apache 2.0; weights are a MiniMax-H3 derivative under the MiniMax H3 Community License.