vdn minimax h3
VDN-Minimax-H3 (VDN-H3) is a hybrid-attention text-to-video model built on MiniMax H3 AI/MiniMax-H3. Its repository provides safetensors weights for a diffusers-based text-to-video pipeline. The model card describes a fast-inference design using a hybrid-attention architecture: one frame-wise linear attention branch plus a softmax branch, along with two small LoRA adapters and a separate linear attention branch that can be merged into the backbone during inference without modifying backbone weights. Reported setup uses SGLang Diffusion on 8xB200 GPUs with 8 denoising steps. The license is the minimax-h3-community-license-agreement, and users are directed to read the license before downloading or running the model. The model card also links to a paper, blog, code repository, weights, and ModelScope page, and notes support for I2VA, FL2VA, and L2VA with the same checkpoint.
At a glance
- Architecture
- unknown
- License
- minimax-h3-community-license-agreement
- Software
- Diffusers, SGLang
Will it run on your machine?
Save your machine to see a personalized rating and its reasoning.
Add my hardwareBest for
- Text-to-video generation
- Fast video generation with a hybrid-attention model
- Inference with SGLang Diffusion on the described GPU setup
- Plug-and-play integration of a separate linear attention branch and LoRA adapters into the MiniMax H3 backbone during inference
- I2VA, FL2VA, and L2VA tasks using the same checkpoint
Tradeoffs
- License is the minimax-h3-community-license-agreement; users must read the license before downloading or running the model.
- The model facts only specify weights and pipeline information; no compatibility, performance, or numerical hardware requirements beyond the reported setup should be inferred.
- Reported fast-inference performance is specific to SGLang Diffusion on 8xB200 GPUs with 8 denoising steps.
- The environment is described as requiring PyTorch 2.13 with CUDA 12.9 and specific package installation steps; other environments are not stated.
- No claims about testing, broader performance, or purchase recommendations are provided in the source data.
Ways to run it
Diffusers · See official guide
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
import torch
from accelerate import cpu_offload_with_hook
from diffusers import ModularPipeline
from diffusers.hooks import apply_group_offloading
pipe = ModularPipeline.from_pretrained("OpenVDN/vdn-minimax-h3", workflow="t2va")
pipe.load_components(trust_remote_code=True, torch_dtype=torch.bfloat16)
apply_group_offloading(pipe.text_encoder, onload_device="cuda", offload_type="leaf_level",
use_stream=True)
_, vae = cpu_offload_with_hook(pipe.vae, execution_device="cuda")
cpu_offload_with_hook(pipe.audio_vae, execution_device="cuda", prev_module_hook=vae)
pipe.transformer.to("cuda")
out = pipe(prompt="a prompt", num_frames=345, num_inference_steps=9,
output=["videos", "audio", "sampling_rate"])SGLang · Linux
Repository-specific command in publisher documentation; confirm dependencies and hardware in the source.
sglang serve --model-path OpenVDN/vdn-minimax-h3 --num-gpus 8 --quantization fp8 --attention-backend hybrid_window_attn_h3 --performance-mode speed --warmup-num-frames 345 --warmup-resolutions 1344x768 --port 30010Explore its uses
Keep exploring
MiniMax H3
An open-weight video model with synchronized native audio. ComfyUI provides a local single-GPU template with dynamic offload; the publisher also documents a separate SGLang multi-GPU route.
Wan-Animate-2 · 14B
Animate a character from a driving video, with text control over the camera viewpoint.
Wan2.1
Explore 2 versions of Wan2.1. Compare model sizes, packages, hardware requirements, and available setups.