H3 Relay · FastH3 VSA Profile

MiniMax H3 is a 33B omni-modal video model, which in plain terms means it is not fast. The standard H3 Relay loaders hand you the full FL2VA/Ref2VA checkpoints and let Sequence Start decide how many steps you burn. FastH3 VSA Profile is the other door: it loads a distilled FastVideo checkpoint that claims to do the whole job in four transformer forwards, and the moment it's in your graph it stops letting you touch the knobs.

Calling it a "loader" undersells it - it's a locked profile. The h3_model output is the exact same H3_RELAY_MODEL wire Generate Shot consumes, but the bundle carries a tag that pins the entire inference contract: Euler sampler, the simple scheduler, exactly four steps, CFG 1 through BasicGuider, video/audio shifts 12/3, Spectrum off, and VSA held at a 10% video-cube keep across the whole denoising range. Your Sequence Start sampler controls and h3_steps still apply to standard bundles, but a FastH3 bundle just overrides them.

How the speed trick works

That 10% keep is the whole story. The checkpoint is FastVideo's step-1300 four-forward VSA release, repacked by Kijai in INT8 ConvRot form, and it was distilled against VSA-H3 running at 90% sparsity. Visual sparse attention that only touches a tenth of the tokens per step is what makes four forwards plausible. That's also why the node fails prompt validation when the VSA runtime isn't present instead of quietly falling back to dense attention - the checkpoint was trained for that sparsity, and a dense fallback would be a different model pretending to be the same one.

The loader itself is thin: it builds a UNETLoader at your chosen weight_dtype, packs the model into the H3 Relay bundle, and computes a cache fingerprint from the model name, dtype, profile, steps, sampler, shifts and keep percentage. Change any of those and you get a fresh, correctly-invalidated artifact.

Inputs and output

• model_name - a single-choice enum pinned to minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors. The node raises ValueError if you hand it anything else; this profile doesn't generalize to other checkpoints.
• weight_dtype - default, or one of the fp8 modes (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) if you're VRAM-capped. The repack is already INT8 ConvRot, so this is your one extra lever.
• manual_cache_revision - per its tooltip, a manual cache reset only. Change it when you replace the model contents without changing the filename; normal model, LoRA, dtype and strength changes are tracked automatically.

Output is a single h3_model (H3_RELAY_MODEL), wired into Generate Shot. The natural partner is Assemble Raw Sequence: FastH3 chains skip the LTX/RIFE finishing stream, so you assemble the raw 480p result directly.

Install - the part that is not "search and click"

This is the node where "install the pack" stops being enough. On top of H3 Relay itself you need three moving pieces the README is explicit about:

• The model, from Kijai's experimental repo:

hf download Kijai/MiniMax-H3-experimental \
  minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors \
  --local-dir models/diffusion_models

• ComfyUI FastVideo-VSA support (upstream PR #15958).
• A VSA-capable comfy-kitchen build (kernel PR #117) and the temporary SolAttnMiniMax patch node attached to that comfy-kitchen PR.

MODELS.md pins the checkpoint's SHA-256, so you can verify the download - worth doing, because this is an experimental preview and FastVideo itself describes it as text-to-audio-video only.

Common issues

• Prompt validation fails at Generate Shot. That's the VSA runtime or the SolAttnMiniMax node missing. Install them; the failure is deliberate, not a bug.
• Adding a LoRA or Attention Backend after the profile is rejected. H3 Relay refuses to apply them to a locked FastH3 bundle because they'd make the profile ambiguous. Remove them, or use a standard loader.
• Thinking it's production-ready. It's not. FL2VA/Ref2VA reference behavior and sliding continuation through this checkpoint are experimental inherited behaviors, not upstream-supported contracts. Keep the standard loaders around as the fallback until a reference-distilled FastH3 release exists - the README says exactly that, and it's good advice.

If you want fast H3 now and can tolerate rough edges, this is the toy to reach for. If you need continuity and quality you can rely on, use the standard profile and let FastH3 mature a few releases.