← NeuralDrift Lab

NEURALDRIFT LAB / TRACKED MODEL · DOCUMENTED

FastVideo FastH3 8-Step V2

FastH3 8-Step V2 is FastVideo's (Hao AI Lab @ UCSD) distilled version of MiniMax-H3, replacing H3's diffusion transformer with a student model trained via data-free DMD2 distillation combined with VSA-H3 sparse attention, cutting inference to 8 transformer forward passes instead of H3's standard range.

Official sources reviewed and requirements verified. No local dependency install or execution attempt has been made — see NeuralDrift Lab status below.

RTX 5080 LAB STATUS

Status: NOT TESTED

NOT TESTED. NeuralDrift reviewed FastH3 8-Step V2's official FastVideo and ComfyUI sources, but has not downloaded its weights, installed either documented path, or attempted a generation on the RTX 5080 Lab machine. The roughly 70GB BF16 number is an arithmetic estimate for the original repository's reported parameter count, not a minimum-VRAM requirement. The official ComfyUI package uses a separately named pruned INT8 diffusion model, but NeuralDrift has not measured its total memory use or verified RTX 5080 compatibility. This status will be updated only after a real NeuralDrift execution.

Relationship to base MiniMax-H3

MiniMax-H3 (MiniMaxAI) is NeuralDrift's existing base model coverage — a text/reference-to-video-with-audio model with native ComfyUI support (Comfy-Org/ComfyUI PR #15224) and official Comfy-Org workflow templates. FastH3 does not replace this coverage; it is tracked separately.

SUPPORTED

What it does

  • Text-to-video with synchronized audio (documented by FastVideo and ComfyUI)
  • Image-to-video with optional first/last-frame conditioning (documented by current official ComfyUI templates)

NOT SUPPORTED IN THIS RELEASE

What it doesn't do

  • Reference-to-video (identity/style/voice lock) — not distilled, per current official ComfyUI documentation

How it's distilled

Method
Data-free DMD2 (Distribution Matching Distillation) student trained against a frozen MiniMax-H3 teacher
Sparsity
VSA-H3 sparse attention, 80% sparsity on video-to-video attention (text/audio stay dense)
Inference steps
8
Scheduler
Video scheduler shift is 10 (base MiniMax-H3 uses 12) — the trained schedule loads from the checkpoint itself, it is not a manual setting to copy from base H3 workflows.

What running it requires

Attention backend
FastVideo's VSA-H3 backend for the upstream reference stack; ComfyUI's BlockSparseAttention node in the official FastH3 templates
Approx. weight size
~70GB (computed, not an official figure)
Native ComfyUI checkpoint?
Yes

Hugging Face metadata reports about 35B BF16 parameters across the original repository. Multiplying that count by 2 bytes estimates roughly 70GB of BF16 weight tensors; this is arithmetic, not an official minimum-VRAM figure or a NeuralDrift measurement. The official ComfyUI package uses a separately named pruned INT8 diffusion model plus shared components, and its documentation does not state a minimum GPU VRAM requirement.

Current official ComfyUI documentation provides FastVideo FastH3 8-Step V2 text-to-video and image-to-video templates, using the Comfy-Org-repacked FastVideo-FastH3-Comfy diffusion model and shared H3 components. This official integration has not been installed or run by NeuralDrift, so no local compatibility, VRAM, generation-time, or A/V-sync conclusion is implied.

Dependencies

  • ComfyUI 0.36.0 or later, the official FastH3 templates, the Comfy-Org-repacked FastH3 diffusion model, and shared MiniMax H3 text encoder and VAEs
  • Alternatively, FastVideo's own reference stack: FastVideo + VSA-H3 via uv using the documented CUDA 13 / Blackwell path
  • FastVideo's tested reference defaults use four NVIDIA B200 GPUs; this is not an RTX 5080 compatibility claim
  • The upstream multi-GPU guidance says the GPU count must divide H3's 56 attention heads; this is irrelevant to a single-GPU setup

GPU compatibility

Distinguishes what NeuralDrift has verified from what upstream or the community claims. Nothing here reflects an actual NeuralDrift execution unless labeled so.

RTX 50-series (Blackwell) — incl. RTX 5080

Unknown

FastVideo's tested reference defaults use four B200 GPUs. The current official ComfyUI FastH3 guide does not make a per-GPU compatibility statement, and NeuralDrift has not installed or run FastH3 on the RTX 5080 Lab machine.

RTX 40-series (Ada Lovelace) / RTX 30-series (Ampere)

Unknown

No official or community report found confirming or ruling out VSA-H3 backend compatibility on non-Blackwell architectures.

AMD / Apple Silicon

Unknown

VSA-H3 is described as a CUDA-compiled backend; no ROCm or Metal/MPS port was found during source review.

MiniMax-H3 vs. FastH3 8-Step V2

Fields NeuralDrift has not measured are marked "Not Tested" or "No Evidence Yet" rather than estimated.

Transformer forward passes

Upstream claim
Base MiniMax-H3
Base H3 schedule
FastH3 8-Step V2
8

Attention sparsity

Upstream claim
Base MiniMax-H3
Dense
FastH3 8-Step V2
80% sparse (VSA-H3, video-to-video only)

Capabilities

Upstream claim
Base MiniMax-H3
Text-to-video, image-to-video, first/last-frame, reference-to-video, audio sync
FastH3 8-Step V2
Text-to-video and image-to-video with optional first/last-frame; no reference-to-video

Generation time (RTX 5080)

Unknown
Base MiniMax-H3
No Evidence Yet
FastH3 8-Step V2
No Evidence Yet

Peak VRAM (RTX 5080)

Unknown
Base MiniMax-H3
No Evidence Yet
FastH3 8-Step V2
No Evidence Yet

Motion quality / fine detail

Upstream claim
Base MiniMax-H3
No Evidence Yet
FastH3 8-Step V2
Below base H3 per the developers' own ablations (see Known limitations)

Audio quality

Upstream claim
Base MiniMax-H3
No Evidence Yet
FastH3 8-Step V2
May be below base H3 per the developers' own disclosure

Setup complexity

Upstream claim
Base MiniMax-H3
Native ComfyUI support, official Comfy-Org templates
FastH3 8-Step V2
Official ComfyUI 0.36+ templates; FastVideo's alternate reference stack uses VSA-H3

RTX 5080 status

Unknown
Base MiniMax-H3
No FastH3 comparison run recorded
FastH3 8-Step V2
Not tested

Known limitations

Upstream claim

The developers explicitly disclose: "Difficult motion, fine detail, and some audio may remain below the base MiniMax H3 model."

Upstream claim

Reference-to-video is not distilled for the official FastH3 ComfyUI integration; use base MiniMax-H3 workflows for reference-based generation.

Unknown

The current official ComfyUI templates have not been installed or executed by NeuralDrift. Their RTX 5080 compatibility, VRAM use, generation time, and A/V sync are unverified.

Unknown

The approximately 70GB BF16 figure is a calculation from upstream parameter metadata, not an official minimum-VRAM requirement or a NeuralDrift measurement.

MiniMax H3 Turbo (Lightx2v + ModelTC)

A separate, unrelated distillation effort — a LoRA applied to the existing base H3 checkpoint rather than a full model swap. Community reporting describes it as running with native ComfyUI support and no custom node required, including an 8-step Ref2VA variant. This is a different project from FastVideo's FastH3 and should not be confused with it.

Source ↗

FastH3 4-Step Preview v1 (community GGUF)

An earlier, separate FastVideo release (not the 8-Step V2 covered on this page) that has a community GGUF quantization and a documented (if involved) ComfyUI path via a custom-compiled attention node.

Source ↗

Sources

WANT TO RUN SOMETHING TODAY?

Local consumer-GPU fit is not yet verified

If you have access to larger VRAM through a cloud GPU provider, that's currently the more realistic path to trying FastVideo's own reference stack. NeuralDrift's cloud GPU hub covers provider options and pricing you can evaluate yourself — treat any pricing there as the provider's own listed rate, not a NeuralDrift benchmark.

Browse cloud GPU providers →

Related on NeuralDrift