Running FastH3 8-Step V2 (MiniMax H3 Distilled) — Current Status

This guide covers FastVideo's FastH3 8-Step V2, a distilled MiniMax-H3 model trained with data-free DMD2 distillation and VSA-H3 sparse attention. Its official FastVideo model card documents a VSA-H3 reference stack, while current official ComfyUI documentation also provides native text-to-video and image-to-video templates for a Comfy-Org-repacked version.

NeuralDrift has reviewed those sources but has not installed or run either path on its RTX 5080 Lab machine. There is no NeuralDrift measurement for FastH3 VRAM, generation time, A/V sync, or RTX 5080 compatibility. The approximately 70GB BF16 figure discussed below is an arithmetic weight-size estimate, not an official minimum-VRAM requirement.

Full sourcing, the GPU compatibility matrix, and the base-H3 comparison table live on the NeuralDrift Lab writeup — this guide is the practical companion to that page.

#Before you start: is this the model you actually want?

There is more than one "fast MiniMax H3" project, and they are easy to confuse:

• →FastH3 8-Step V2 (this guide) — built by FastVideo / Hao AI Lab @ UCSD. The original FastVideo card documents text-to-audio-video with its VSA-H3 reference backend. Current official ComfyUI documentation separately provides native text-to-video and image-to-video templates with optional first/last-frame conditioning. Neither path has been run by NeuralDrift.
• →MiniMax H3 Turbo — a separate project by the Lightx2v and ModelTC teams, distributed as a LoRA on top of the existing base H3 checkpoint. Community reporting describes it running with native ComfyUI support and no custom node required, including an 8-step reference-to-video variant. NeuralDrift has not tested this either, but it is architecturally a much smaller lift than FastH3 8-Step V2 — see Related projects below.

If your goal is "the fastest thing I can actually get running in ComfyUI today," the Turbo LoRA is the more realistic starting point. If your goal is specifically FastVideo's FastH3 release, keep reading.

#Current official integration paths

There are two current upstream-documented paths:

• →FastVideo reference stack: The official model card requires FastVideo's VSA-H3 attention backend and documents installation through uv using its CUDA 13 / Blackwell path. Its tested defaults use four B200 datacenter GPUs. That is reference-environment documentation, not a consumer-RTX compatibility result.
• →Native ComfyUI templates: Current official ComfyUI FastH3 documentation lists ComfyUI 0.36.0+, a Comfy-Org-repacked FastVideo-FastH3-Comfy diffusion model, shared MiniMax H3 components, and text-to-video/image-to-video workflow templates. It does not state a minimum GPU VRAM figure or make an RTX 5080 compatibility claim.

Hugging Face metadata for the original FastVideo repository reports roughly 35B BF16 parameters. At two bytes per BF16 parameter, that is about 70GB of weight tensors. This is a computed size estimate across the original repository's parameter metadata, not an official minimum-VRAM specification, a ComfyUI package-size statement, or a measured NeuralDrift result. The ComfyUI integration distributes a separately named pruned INT8 diffusion model, so its runtime memory requirement cannot be inferred from the BF16 arithmetic alone.

#Current capability documentation

The original FastVideo card describes text-to-audio-video and says FL2VA and Ref2VA were not distilled. The current official ComfyUI guide for its FastH3 integration describes text-to-video and image-to-video templates, including optional first/last-frame conditioning; it continues to exclude Ref2VA. NeuralDrift has not run either integration, so this is source documentation, not an independently verified capability or A/V-sync result.

#What doesn't exist yet

The earlier, separate FastH3 4-Step Preview v1 has a community GGUF quantization (realrebelai/FastH3_GGUFs). That community project is for the 4-step model, not this 8-Step V2 release. For the current 8-Step V2 release, use the official ComfyUI templates or FastVideo reference instructions above. NeuralDrift has not independently verified their consumer-GPU memory behavior.

#If you have the hardware to actually try it

For the official ComfyUI route, begin with the FastVideo FastH3 workflow examples. For the FastVideo reference route, begin with its own repository and model card:

• →FastVideo/FastVideo-FastH3-8-Step-V2 on Hugging Face
• →Hao AI Lab's FastH3 technical writeup

Follow the documented path that matches your route. NeuralDrift cannot yet recommend either path for a specific consumer GPU, because it has not installed or executed FastH3 locally.

#What to run today instead

Base MiniMax-H3 already has native ComfyUI support and official Comfy-Org templates, and NeuralDrift's Lab machine already has the Ref2VA checkpoint installed locally. It supports reference-to-video, which current FastH3 documentation excludes. Start with the official ComfyUI MiniMax H3 guide if you want synchronized video+audio generation working locally today.

If you specifically want a local NVIDIA video+audio workflow that NeuralDrift has actually execution-tested end to end, the catalog's AnimateDiff and ACE-Step 1.5 audio workflows have current, successful RTX 5080 records — see NeuralDrift Lab for the full evidence index.

#No local GPU with enough headroom?

NeuralDrift's cloud GPU hub lists provider options if you want to evaluate larger-VRAM instances yourself. Treat any pricing shown there as the provider's own listed rate, not a NeuralDrift benchmark.

#Related projects

See the full writeup on the NeuralDrift Lab page for FastH3 8-Step V2 for the complete source list, the GPU compatibility matrix (tested vs. upstream claim vs. community reported vs. unknown), and the field-by-field comparison against base MiniMax-H3.

#NeuralDrift Lab status

Not tested. NeuralDrift has reviewed FastH3 8-Step V2's official sources but has not downloaded its weights, installed either documented integration, or attempted a generation. This guide will be updated with real execution results only if that changes.