01
What is VDN-H3?
VDN-H3, short for Video DeltaNet on MiniMax H3, is an independent OpenVDN research release built from MiniMax H3. The project adds a hybrid-attention path rather than replacing the whole backbone: nearby temporal context stays on softmax attention while distant context is handled by frame-wise linear Video Delta Attention.
OpenVDN released the project page, model weights, optimized inference stack and training code on September 6, 2026. On September 8, the project README added image-to-video and first/last-frame-to-video support using the same checkpoints; the multi-GPU Ulysses scripts remain documented as text-to-video only.
02
What OpenVDN released
The official Hugging Face bundle is about 82 GB: roughly 72 GB of released MiniMax H3 base components, a 4.3 GB 50-step VDN stage and a 5.1 GB distilled 8-step stage.
- 50-step checkpoint: stage-b-step-2000, containing the learned linear branch and default LoRA adapters
- 8-step checkpoint: stage-dmd-step-250, adding the matching turbo adapter used for the few-step path
- training and inference code in the OpenVDN GitHub repository
- single-GPU and dedicated eight-GPU H200/B200 inference scripts
03
How fast is VDN-H3?
OpenVDN reports 11.23 seconds of steady-state denoising for a 14.4-second, 768p clip using the 8-step checkpoint on eight NVIDIA B200 GPUs. That is faster than playback for the denoising stage.
The same official table reports about 5.3 minutes at 50 steps and 51 seconds at 8 steps on one B200; about 1.2 minutes and 11.23 seconds respectively on eight B200s; about 9.4 minutes and 90.5 seconds on one H200; and about 1.9 minutes and 18.3 seconds on eight H200s.
04
What does “near-lossless quality” mean here?
OpenVDN describes VDN-H3 as maintaining near-lossless visual quality relative to dense H3 and publishes qualitative comparisons. We did not find an independent standardized quality benchmark that numerically proves that phrase, so it should remain an author claim rather than a third-party finding.
05
8-step vs 50-step VDN-H3
The 50-step stage is useful when you want to evaluate the hybrid-attention architecture with a higher-step generation path. The 8-step DMD stage adds few-step distillation and is the checkpoint behind OpenVDN's fastest published numbers.
If practical generation speed is the goal, the 8-step checkpoint is the obvious starting point. If you are studying the architecture itself, the 50-step route better separates the attention change from the additional speed gained through distillation.
06
Can you run VDN-H3 locally?
Yes. The current official setup uses Python 3.12, recommends PyTorch 2.13 with CUDA 12.9, requires the FlashAttention 4 path and installs a patched Diffusers build through the project setup script. The first render compiles kernels and can take several minutes before later runs reuse the cache.
OpenVDN provides a generic single-GPU script plus dedicated eight-GPU scripts for H200 and B200 hardware. It does not publish one universal consumer-GPU minimum-VRAM requirement, so community RTX results should not be presented as official hardware requirements.
07
Text, image and keyframe conditioning
The quick-start text-to-video path encodes a prompt with Qwen3-VL-32B before rendering through the diffusion model. The September 8 update also documents image-to-video and first/last-keyframe conditioning with the same released checkpoints.
For first/last-frame generation, the official encoder places the images on a 768-short-edge canvas, processes them with Qwen3-VL and VAE-encodes the conditioning latents into the prompt file. OpenVDN currently documents that keyframe-conditioning route on the single-GPU entrypoint, not the multi-GPU Ulysses scripts.
08
Code and model weights use different licenses
OpenVDN's training and inference code is licensed under Apache 2.0. The VDN-H3 weights are a MiniMax H3 derivative and are distributed separately under the MiniMax H3 Community License Agreement.
09
Is there a hosted VDN-H3 API?
The official OpenVDN release is primarily weights plus code. A Hugging Face Space exists for demonstration, but we did not verify a generally available official hosted API with published VDN-H3 pricing or quotas. Do not infer API availability from a demo Space.
10
Important limitations
- the fastest published result needs eight B200 GPUs
- headline timing excludes several end-to-end pipeline stages
- near-lossless quality is an OpenVDN claim supported mainly by qualitative comparisons
- the official stack is datacenter-oriented and does not define a universal consumer minimum VRAM
- weight licensing has territorial and redistribution restrictions
- hosted API pricing and quotas were not verified
11
Bottom line
VDN-H3 is notable because it combines a hybrid attention change with an 8-step distilled path and publishes both the model artifacts and the training/inference stack. The most useful next step for developers is the official local-running guide below, which keeps the H200/B200 benchmark claims separate from unverified consumer-hardware assumptions.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.