Video Rebirth releases HyperFlow H3 Flash, a 14B few-step video and audio model compressed from MiniMax-H3
Video Rebirth has published HyperFlow H3 Flash, a compressed, few-step version of MiniMax-H3 with 13.7B to 13.9B parameters. It generates 5.2 to 15 second videos at 24 fps with synchronized 32 kHz stereo audio, using 4, 6 or 8 sampling steps without classifier-free guidance.
Video Rebirth has released HyperFlow H3 Flash, a 14B-class video model that generates video and synchronized stereo audio together in as few as 4 sampling steps. The weights are on Hugging Face under VideoRebirthAI/hyperflow-h3-flash. The model card describes it as Video Rebirth's compressed, few-step version of MiniMax-H3, and states that the repository is not an official MiniMax product.
Specifications
- Parameters: 13.7B (fl2va) or 13.9B (ref2va), less than half of the ~31B MiniMax-H3 transformer, per the model card.
- Output: 5.2 to 15 s video at 24 fps (1344×768 by default) with a 32 kHz stereo soundtrack.
- Sampling: 4, 6 or 8 steps without classifier-free guidance (CFG). The default is 6 steps.
- Tasks: t2va (text prompt), fl2va (prompt plus a first keyframe, or first and last keyframes), and ref2va (prompt plus up to 12 references).
- Context window and pricing: not applicable to this open-weights release; no hosted pricing is listed.
Speed
For a 5.2 s clip at 6 steps, the model card lists 8.7 s on 4× B200, 36 s on 1× H200, and 92 s on an RTX 5090. It also says NVIDIA SOL-H3, which it calls its inference-acceleration partner, runs the model in real time on 4× B-series GPUs and supports it from day 0. These are the developer's figures.
Weight formats and hardware
The repository stores only the compressed transformers. Text encoder, VAEs, tokenizer and processor load from upstream repositories, and the Python package never downloads the 62 GB base transformer.
- GGUF: six quant levels from Q3_K_M to Q8_0 (7.2 to 15.2 GB). Q4_K_M is the recommended quant at 9.1 GB (fl2va) and 9.2 GB (ref2va). Q8_0 is 14.9 GB and 15.2 GB. Runs in ComfyUI, Unsloth Studio and stable-diffusion.cpp.
- INT8 convrot: W8A8 INT8 plus low-rank, about 16 GB. Aimed at H100, H200, B200 and B300 GPUs (1 to 8), or 24 GB+ RTX GPUs. ComfyUI single-file copies are 16.0 GB (fl2va) and 16.4 GB (ref2va).
The card says the model runs on 24 GB and 32 GB GPUs in PCs with 16 GB of system RAM. 16 GB GPUs are not measured yet. The Python package needs Python 3.10+, PyTorch 2.6+ with CUDA and Triton, an Ampere or newer NVIDIA GPU, and diffusers 0.40 or later. First run downloads 90 GB (68 GB with the fp8 text encoder). Multi-GPU use requires Linux or WSL2.
GGUF supports t2va and fl2va, but for ref2va, GGUF and ComfyUI INT8 accept images only. The Python INT8 package accepts images, videos and audio for ref2va.
License
Use is governed by the MiniMax H3 Community License Agreement, which includes an Acceptable Use Policy. It restricts territory: the weights are not licensed in the EU, UK, South Korea, or the US without prior authorization from MiniMax.
What this means
The release targets the cost of running a joint video and audio model. Fewer steps without CFG, a parameter count under half the base transformer, and GGUF and INT8 builds put the model on 24 GB and 32 GB consumer GPUs, according to the developer's measurements. Whether 16 GB cards can run it is still open. The territory restrictions in the license matter for anyone planning deployment in the listed regions, and require prior authorization from MiniMax in each case. Video Rebirth also points to an 8-step LoRA for the full-size MiniMax-H3 transformer for users who want to stay on the larger model.
Related Articles
PrunaAI releases P-H3-Image: 8 MP image generation and editing in 8 steps, built on MiniMax-H3
PrunaAI has released P-H3-Image, an image model based on MiniMax-H3 that renders 8-megapixel images in 8 steps with no classifier-free guidance. It handles text-to-image and editing with up to 8 input images, and bf16 inference needs about 101 GB of GPU memory.
OpenCSG Agentic-27B: 27.36B-Parameter Tool-Use Model Runs on One A800, Commercial Use Needs Permission
OpenCSG has released Agentic-27B, a 27.36B-parameter model for multi-step tool use, built on the dense Qwen3.8-27B. The company reports it runs on a single 80GB NVIDIA A800, but commercial use requires authorization under the OpenCSG license, and all benchmark results are OpenCSG's own.
ByteDance Seed releases Cola2, a 9B continuous-latent language model with 60.10 eight-task average
ByteDance Seed has published Cola2, a 9B-parameter base language model that generates continuous latent blocks with a diffusion transformer and realizes them as discrete tokens. Trained on 346.03B source tokens, it reports a 60.10 average across eight tasks. Weights are on Hugging Face; training code is not included.
ByteDance's Seedream 5.5 Holds Up Through 100 Edit Rounds in QbitAI Test; 2K Edit Costs About 0.28 Yuan
ByteDance's Seedream 5.5 stayed more faithful than GPT-Image 2.6 in QbitAI's 100-round sequential editing test, though some distortion appeared in later rounds. According to official information, a 2K square image edit via API costs about 0.28 yuan, under half the price of Seedream 5.0 Pro.
Comments
Loading...