AI Briefing
KOSign in

Video Rebirth Releases HyperFlow H3 Flash, a 13.7B Compressed Video-Audio Diffusion Model

·2026.11.08 00:16

Key point

The model compresses MiniMax-H3 to 13.7B parameters for 4-8 step generation, running on consumer RTX 5090 GPUs in 92 seconds for 5.2-second clips.

Details

  • Video Rebirth released HyperFlow H3 Flash, a compressed version of the 31B MiniMax-H3 model that reduces parameters to 13.7B (fl2va) or 13.9B (ref2va) while supporting 4, 6, or 8 sampling steps without CFG.
  • The model generates 24 fps video at 1344x768 resolution with 32 kHz stereo audio, achieving 92-second generation for 5.2-second clips on a single RTX 5090 and 8.7 seconds on 4x B200 GPUs.
  • It supports three tasks: text-to-video-audio (t2va), first/last keyframe conditioning (fl2va), and reference-based generation (ref2va) using up to 12 references (9 images, 3 videos, 3 audio files).
  • Two runtimes are provided: GGUF for consumer PCs with 64 GB system RAM (supporting ComfyUI and stable-diffusion.cpp) and INT8 convrot for data-center GPUs or high-RAM systems (supporting Python and diffusers).
  • Minimum requirements include 16 GB system RAM and 24/32 GB GPU VRAM for Q4_K_M (9.1 GB) or Q8_0 (14.9 GB) quantizations, with INT8 transformers running 10-20% faster than GGUF but using more system RAM.
  • The model uses structural pruning, low-rank decomposition, and rotated W8A8 INT8 quantization, with a two-time (t, r) conditioning mechanism enabling low-step sampling.
  • License is Apache 2.0 for code, but model weights are subject to the MiniMax H3 Community License Agreement, restricting use in the EU, UK, Korea, and US without MiniMax approval.
open-sourceai-infrastructurequantizationvideo-generationmodel-compressiondiffusion-models
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Google Cloud AIGoogle Cloud announces general availability of Gemini Omni 1.2 Flash for video generationKakaoKakao Releases Open-Source 3B Multimodal Decision-Making Model 'Kanana-2-v-decide'Hacker NewsAnthropic amended S-1 reveals $17 billion Q3 revenue and $895.4 billion in compute commitmentsRedditOpen-source EOU detector for Indic languages achieves 91.3% TPR on LiveKit AgentsRedditRicursive Intelligence releases 65B dense model Ripple-65B Preview under Apache 2.0GeekNewsDevvm: Open-Source Tool Creates Isolated VMs for AI Coding Agents