Show HN: open-source compression of a 31B video+audio diffusion model to 14B, 6 steps, INT8
jiant · hn · 2026-11-07
An open-source project, hyperflow-h3-flash, compresses a 31B-parameter video+audio diffusion model down to 14B with 6 sampling steps and INT8 quantization, substantially cutting inference cost for multimodal generation.
More from Multimodal
- Fable 5.2 recreates LOTR scene with self-built engine and 6cm FLIP water sim — AymericRoucher · 2026-11-11
- Game character design sheet made with MiniMax-H3-Image shows cross-view consistency — eyishazyer · 2026-11-11
- MiniMax H3 open-weights image-to-video nails lip sync for three speakers in user test — BarnacleElectronic80 · 2026-11-11
- AirCad's Built-in Agent 3D-Prints Lab Equipment Designs in ~30 Seconds — iskander · 2026-11-11
- WanGP v19.28: 16MP Images in 6 Steps, 2K Video Regeneration and More — cocktailpeanut · 2026-11-11
- Phosphene: free open-source local AI video on Mac with Hailuo H3 and LTX-2.5 — cocktailpeanut · 2026-11-11