Is HyperFlow H3 Flash state of the art?Pending benchmarks.Announced Nov 6, 2026; no independent results yet.
Video Rebirth's compressed few-step MiniMax-H3 derivative for joint video and synchronized stereo audio (not an official MiniMax product). About 14B parameters (13.7B fl2va / 13.9B ref2va), default 6 sampling steps without CFG, with INT8 convrot and GGUF packs on Hugging Face. Supports text-to-video+audio, first/last-frame conditioning and reference-to-video. Weights are a Model Derivative under the MiniMax H3 Community License (territorial restrictions apply).
The details so far
- Lab
- Video Rebirth
- Announced
- Nov 6, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 14B (13.7B fl2va / 13.9B ref2va)
- Role
- Few-step joint video and stereo audio generation (t2va, fl2va, ref2va)
- Base
- Compressed / distilled from MiniMaxAI/MiniMax-H3; separate from the MiniMax H3 catalog page and other H3 derivatives
- Steps
- 4, 6 or 8 steps without classifier-free guidance; default 6
- Output
- 5.2–15 s video at 24 fps (1344×768 default) with 32 kHz stereo soundtrack
- Formats
- INT8 convrot (~16 GB) and GGUF Q3_K_M–Q8_0 (7.2–15.2 GB) for fl2va and ref2va partitions
- Checkpoints
- VideoRebirthAI/hyperflow-h3-flash on Hugging Face; code at github.com/Video-Rebirth/hyperflow-h3-flash
- License
- MiniMax H3 Community License Agreement (HF license tag: other / license_name minimax-h3-community-license-agreement). Territory limits (not licensed in the EU, UK, South Korea, or the US without prior MiniMax authorization) and an Acceptable Use Policy apply
- Related
- Shipper is Video Rebirth; MiniMax H3 is the base. Separate from the older HyperFlow 8-step LoRA on full-size MiniMax-H3 (VideoRebirthAI/hyperflow)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: end-to-end 5.2 s clip at 6 steps takes 8.7 s on 4× B200, 36 s on 1× H200, and 92 s on an RTX 5090 (model card).
Lab claim: Q4_K_M GGUF is 9.1 GB and runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM (16 GB GPUs not measured).
Lab claim: structural pruning drops 3 of 50 fl2va blocks (1 of 50 for ref2va) and factors 102 / 114 linear layers, with rotated W8A8 INT8 on 180 layers.
Sources
REPOVideoRebirthAI/hyperflow-h3-flash model card (Hugging Face)huggingface.co/VideoRebirthAI/hyperflow-h3-flashREPOVideo-Rebirth/hyperflow-h3-flash (GitHub)github.com/Video-Rebirth/hyperflow-h3-flashLABVideo Rebirth on X: HyperFlow H3 Flashx.com/video_rebirth/status/2118729708988969078 · Nov 6, 2026LABVideo Rebirth on X: HyperFlow H3 Flash weightsx.com/video_rebirth/status/2118729711086076135 · Nov 6, 2026Tracking continues until an independent board lists it.