MiniMax H3 Add-Ons Are Piling Up on Hugging Face: 8-Step LoRAs, a WASD World Model & More

MiniMax H3's open weights logged 4,299,737 downloads in the past month on Hugging Face as of September 19, 2026.

ComfyUI's repackaged copy of the same model shows another 20,497,092 over the same window. In September, outside teams shipped speed-ups, a streaming runtime, a keyboard-driven world model and reusable reference files, all built on MiniMax's H3.

The newest is HyperFlow from Video Rebirth, whose code repository went up on September 17. HyperFlow is an 8-step LoRA: Diffusers' default schedule runs 49 model forwards and HyperFlow runs 8, per its model card. Video Rebirth reports about a 3x end-to-end gain, roughly 60 seconds against 175 seconds per clip on four H200 GPUs. NVIDIA's research repository merged HyperFlow support on September 19.

ComfyUI sped up the other end of the pipeline. Version 0.36.0, published September 15, includes developer kijai's H3 VAE optimizations. The kernel pull request reports fp16 encoding falling from 46.6 to 32.0 seconds and int8 decoding from 18.7 to 12.9 seconds on an RTX 5090, at 1344x768 over 362 frames.

Alibaba's TaoLive AIGC team posted TaoMate-H3, a 3-step streaming LoRA and runtime, with a news entry dated September 8. On eight H20 96 GB GPUs, the team measured 14.8 seconds of pure transformer time for a 10-second 480x864 clip, against 169.6 seconds for stock H3.

H3-World turns H3 into an interactive world model. Authors from Tencent, the National University of Singapore and The Hong Kong Polytechnic University trained a 65.6M-parameter LoRA on 8,000 gameplay clips. W, A, S and D move the character; I, J, K and L move the camera. The default run produces a 5.2-second clip at 832x480, per the code repository.

RefMods tackle character consistency. The ComfyUI-MiniMaxH3Mod node pack saves image, video or audio references as reusable .safetensors files with no training run. Fizgig 6.0.1 adds a version tuned against H3 itself.

What Is MiniMax H3?

MiniMax H3 is MiniMax's video model with native stereo audio, launched on July 31, 2026. The weights followed on August 3, according to RunPod. The model card describes a 33B-parameter transformer generating 4 to 15 seconds at 24 fps.

Parameter counts vary by page. ComfyUI Wiki lists a 22.1B backbone for TaoMate-H3, a figure that appears on the TaoMate paper's project page. MiniMax's card says 33B, with about 13B in branches that need not be loaded for inference.

Now, the catches. Local output is 768p; 2K needs MiniMax's hosted H3-Regenerate-2K stage. HyperFlow's card cites peak memory of about 80 GB per card, and TaoMate-H3's validated setup is eight data-centre GPUs. HyperFlow and H3-World both warn that a stock LoRA loader will not reproduce their results. The RefMod README concedes that identity, clothing and background can still mix.

Licensing follows the base model. H3-World's code is Apache 2.0, but its checkpoint falls under MiniMax's licence. The H3 licence also excludes the EU, UK, South Korea and US, as SLOP TV's explainer covers.

More is scheduled. TaoMate-H3's first-and-last-frame version is due before October 15, 2026, per its card. MiniMax says its sparse-attention code and the 2K module will be released later, with no date given.

MiniMax H3's weights are on Hugging Face, and ComfyUI users get the VAE speed-up by updating to 0.36.0 or later. The original HyperFlow file is not a drop-in ComfyUI LoRA; ComfyUI users need a community conversion or the ComfyUI-HyperFlow node pack, according to ComfyUI Wiki.