MiniMax H3 Workflow Advisor

Every H3 launch thread has the same comment, buried under the hype: "I've got a 12GB card, can I even run this?" The MiniMax H3 Workflow Advisor is the answer node - a five-field form that tells you what to expect from your hardware before you burn an afternoon downloading 42.5GB of weights that won't fit comfortably. It's a rule-based sanity checker, not a benchmark, and it's careful to say so: every output starts by labeling itself "practical starting guidance, not an official hardware minimum."

That disclaimer isn't modesty. H3 is new enough that nobody has a verified hardware floor, so this node encodes community-shaped wisdom instead of official numbers. The real-world picture, straight from the people running it: a 3060 technically runs it but "very slowly," 12GB-class cards mean aggressive offloading, and a 16GB card can still blow past VRAM at the text-encoder or VAE-decode stage even when sampling starts fine.

How it works

You tell it your mode (T2VA, I2VA, FL2VA, REF2VA), your vram_gb (4–192, default 16), your system_ram_gb (8–1024, default 64), the resolution you're after (Lower than 768p / 768p class / 1080p / 2K), and your duration_seconds (4–30). It applies a simple tiered model:

• ≤12GB VRAM - treated as experimental; expect aggressive offloading and a recommendation to reproduce on your exact stack before trusting it
• ≤16GB - validate a short, lower-resolution run first; explicit warning that text encoder and VAE decode can OOM even when sampling succeeds
• ≤24GB - you can use the native workflow, but check memory headroom before pushing output size
• >24GB - benchmark one variable at a time instead of changing everything at once

Then it layers on the subtle stuff people skip: under 32GB system RAM your CPU offload is constrained; under 64GB, free RAM isn't the same as usable offload capacity. A 2K target triggers a reminder to verify whether your workflow is a true 2K path or just an upscale, and anything past 10 seconds flags that memory and time rise materially with duration. It even ends with a pragmatic fallback: if the local setup costs more time than the job is worth, compare with a browser-based run.

The input that matters most

vram_gb is the whole ballgame - the advice tiers are built around it. Set it honestly, because the node can't see your card. system_ram_gb is the one people underestimate: H3's offloading path leans on it hard, and a beefy GPU with 16GB of system RAM is a worse setup than the numbers suggest.

The single advice output is a plain-text block you read, copy, or wire into a text viewer. It's not a loader, not a workflow generator - it's the "stop, think, then start small" note your past self would appreciate.

Install

ComfyUI Manager, search MiniMax H3 Workflow Toolkit, or:

cd ComfyUI/custom_nodes
git clone https://github.com/jayseanbrambila-del666/ComfyUI-MiniMax3-H3-Toolkit

Restart and search "MiniMax H3". No third-party dependencies, no models, no network calls.

Common gotchas

Don't read the output as gospel. It's tiered guidance, and a 12GB card can run H3 - the node isn't saying no, it's saying "expect pain, start small, don't trust a single successful run." Pair it with the H3 Prompt Builder and Run Summary for the full loop: advice tells you where to start, the builder writes the prompt, the summary logs what you ran.

One reminder that isn't a gotcha but matters: the H3 weights themselves are geofenced by the Community License out of the US, EU, UK, and South Korea. The Advisor will cheerfully give you advice anywhere - whether you're licensed to run the weights where you are is a separate question it doesn't answer.