HyperFlow H3 Flash: 13.7B Few-Step Model Generates Video With Stereo Audio, Runs on RTX 5090
Apprehensive_Sky892 · reddit · 2026-11-07
- Video Rebirth released HyperFlow H3 Flash, a compressed few-step version of MiniMax-H3 (31B) at 13.7B (fl2va) / 13.9B (ref2va) params.
- Generates video plus synchronized 32 kHz stereo audio: 5.2–15 s at 24 fps, default 1344×768.
- Only 4/6/8 sampling steps, no CFG. A 5.2 s clip at 6 steps takes 8.7 s on 4× B200, 36 s on one H200, 92 s on an RTX 5090.
- Real-time inference on 4× Blackwell GPUs via NVIDIA's SOL-H3 acceleration engine.
- Q4KM quantized size is just 9.1 GB; runs on 24/32 GB GPUs with 16 GB system RAM.
- Ships with a ComfyUI custom node (needs ComfyUI-GGUF), plus open-source code, setup guides and workflows on GitHub/Hugging Face.
More from Infra
- vLLM contributor: lossless KV cache quantization viable but not worth it — AlpinDale · 2026-11-07
- NVIDIA's first co-packaged optics Ethernet switch, Spectrum-X Photonics, enters full production — TinfoilTricorn · 2026-11-07
- This week's open training stack: MoE training handbook, harness adapters, and a harness leaderboard — ariG23498 · 2026-11-07
- Musk: space will be the cheapest place to run superintelligence within 2-3 years — TinfoilTricorn · 2026-11-07
- JITEngine synthesizes custom LLM inference engines for Apple Silicon, claims up to 2.7x speedup — AccBalanced · 2026-11-07
- Dev predicts 2027's local inference boom: integrated hardware + OS + inference stack, not just GPUs — vincent_koc · 2026-11-07