MiniMax H3 video ecosystem grows: OpenVDN's VDN-H3 (trending #12) and community fine-tune Minimax-h3_Singularity (trending #22) hit Hugging Face trending

Two distinct MiniMax H3-based video models trended on Hugging Face within four days: OpenVDN's VDN-H3, a hybrid-attention add-on that renders 768p 14.4-second clips in 11.23 seconds on 8 B200 GPUs, and WarmBloodAban's Minimax-h3_Singularity, a community…

Two separate MiniMax H3-derived video generation models appeared on the Hugging Face trending chart between 2026-09-02 and 2026-09-05. On 2026-09-02, OpenVDN's VDN-Minimax-H3 (VDN-H3) trended at #12: per the report, it adds a frame-wise linear attention branch plus two LoRA adapters to MiniMax H3, distilled into 8-step and 50-step variants, and renders a 768p, 14.4-second clip in 11.23 seconds on 8 B200 GPUs (90.5 seconds on a single H200) using 8 denoising steps. Weights total about 82 GB including the 72 GB H3 base, and the release is under the MiniMax H3 Community License, which per the report excludes the EU, UK, Korea, and the US. On 2026-09-05, WarmBloodAban's Minimax-h3_Singularity trended at #22: a community fusion fine-tune built from multiple checkpoints with pruning and weight optimization, supporting Text-to-Video, Image-to-Video, Reference-to-Video, and Video-to-Video workflows in ComfyUI, and claiming improvements in HDR clarity, distant face restoration, motion fluidity, and fantasy VFX; it recommends the minimax_h3_ref2v_turbo_4step_v0.1 LoRA for four-step accelerated inference and offers a RunningHub demo. The two reports describe independent releases and contain no conflicting facts.

• VDN-H3 (OpenVDN/vdn-minimax-h3) trended #12 on Hugging Face as of 2026-09-02T05:03:41Z.
• Minimax-h3_Singularity (WarmBloodAban/Minimax-h3_Singularity) trended #22 on Hugging Face as of 2026-09-05T18:51:01Z.
• VDN-H3 speed (per Report 1): 768p, 14.4-second clip in 11.23 seconds on 8 B200 GPUs; 90.5 seconds on one H200; 8 denoising steps.
• VDN-H3 architecture (per Report 1): frame-wise linear attention branch plus two LoRA adapters merged at inference without touching backbone weights; 8-step variant distilled from a 50-step model; distributed FP8 inference on B200s and…
• VDN-H3 weights total about 82 GB, including the 72 GB H3 base; prompts are encoded with Qwen3-VL-32B; H3-Context-IR rewriting is recommended for quality; inference path built on Diffusers, FlashAttention, and Triton.
• VDN-H3 licensing: MiniMax H3 Community License, which per Report 1 excludes the EU, UK, Korea, and the US.
• Minimax-h3_Singularity is a fusion fine-tune of MiniMax-H3 built from multiple checkpoints with pruning and weight optimization, supporting T2V, I2V, Ref2V, and V2V in ComfyUI.
• Singularity's claimed improvements (per Report 2): HDR clarity, distant face restoration, de-glossed aesthetics, enhanced combat motion, and fantasy VFX.