MiniMax H3 Community Ships 11 More Projects: 8-Step Generation, TPU Support, and Slimmer Reference Tokens

MiniMax H3 is an open-source video generation model with native stereo audio and multimodal reference control. The official team has released a new batch of community contributions, adding 11 entries to the ecosystem index.

Several directions in this update are worth noting.

Fewer steps

• FastH3 Omni 8-step reference-conditioned video + audio: supports up to 9 images, 3 videos, and 3 audio inputs, and outputs 24fps and synchronized 32kHz stereo.
• FastH3 4-step preview and LightX2V Turbo LoRA: reduce sampling to 4 steps.
• PDD 8-step Acc LoRA: implemented by Alibaba PAI based on NVIDIA's distillation method, and loads natively in ComfyUI.

Stronger attention optimization

VDN (Video Delta Net) changes the architecture: sliding-window softmax + frame-level linear attention. A 768p, 14.4-second clip takes 90.5 seconds to denoise on a single H200, and 11.23 seconds on 8×B200—faster than real-time playback.

NVIDIA's Sol-Attn is 1.14–1.44× faster than SageAttention on RTX 5090, with 37% lower peak VRAM.

Reference token compression

VPack compresses Ref2VA video reference tokens and is independent of other acceleration methods.

More runtime environments

• antirez's h3.c: a native Metal engine for Apple Silicon, MIT licensed.
• TorchTPU: runs H3 video/audio examples on Cloud TPU v6e-4.
• FastH3 Trim NVFP4: for Blackwell hardware such as DGX Spark.

The value of this work lies not in any single number but in the combination. Cheaper experiments, finer control, and more options—GPU, TPU, and Mac can all run it. Open weights let the community modify methods, tune runtimes, and build workflows. For an open-source model, this is the real expansion of its ecosystem niche.

The full index is on GitHub: awesome-minimax-h3-integration

Thanks to everyone who trained, quantized, tested, and shared.