TL;DR
| Local (ComfyUI) | Online (minimaxh3.co) | |
|---|---|---|
| Resolution | 768p (2K needs the hosted regeneration pass anyway) | Up to 2K |
| Cost structure | Hardware you already own or buy, ~free electricity | Free starter credits, then $9.90–$99.90 credit packs |
| Setup time | Real — ComfyUI, weights, quantization decisions | None — browser only |
| Speed per clip | ~6 min for 20s at 608×352 on an RTX 4090 | Seconds to a few minutes, no local compute needed |
| Quality control | You review every clip yourself | Same underlying model — same output caveats apply |
| License restrictions | US/EU/UK/Korea excluded — see License | Governed by the hosted product's own terms |
| Best for | Volume, privacy, GPU already owned, tinkering | One-off use, no GPU, need 2K, in a restricted region |
The Real Cost Comparison
Local — What You're Actually Paying
There's no license fee for local generation, but "free" undersells three real costs:
- Hardware. Our tested setup used a single RTX 4090 (~46GB VRAM peak) — a $1,600–2,000 card if you're buying one specifically for this. If you already own a capable GPU for other work, this cost is effectively zero.
- Electricity. Genuinely small — a 20-second clip at ~6 minutes of generation time on a ~450W card works out to roughly 45Wh, a fraction of a cent at typical residential electricity rates. Not a real factor in the decision.
- Your time. Setup (ComfyUI, model downloads, quantization choices — see ComfyUI setup) is a one-time cost. The ongoing cost is quality-checking every clip yourself — listening to audio, checking hands/objects/text — since our testing found real, recurring issues that need human review.
Online — What You're Actually Paying
MiniMax H3 online uses one-time credit packs, not a subscription:
| Plan | Price | Credits | Rate |
|---|---|---|---|
| Free starter | $0 | Starter credits | 720p, watermarked, personal use only |
| Starter | $9.90 | 99 credits | $0.10/credit, commercial-use license |
| Basic | $29.90 | 370 credits | $0.08/credit, priority queue |
| Plus | $49.90 | 700 credits | $0.07/credit, up to 5 concurrent jobs |
| Professional | $99.90 | 1,665 credits | $0.06/credit, fastest queue |
2K renders use more credits per second than lower-resolution drafts — see the pricing page for the current exact rate before estimating a monthly cost.
When Each Actually Wins on Cost
- You already own a 24GB-class GPU and generate a lot of video → local is genuinely near-free after setup time.
- You generate occasionally, or don't own a capable GPU → buying credits is cheaper than buying a GPU you'd rarely use.
- You need 2K specifically → the open weights alone cap at 768p; reaching 2K still routes through MiniMax's hosted regeneration pass either way, which narrows the local cost advantage for that specific output.
The Real Quality Comparison
Resolution — 768p Local vs 2K Online
This is the most-misunderstood fact from the open-source hub page: the open weights generate at 768p. The model's headline 2K output specifically requires H3-Regenerate-2K, which currently only runs on MiniMax's hosted infrastructure — not something the local open-weight path produces on its own.
Reliability — Same Model, Same Underlying Limits
Our 50-clip test on a local RTX 4090 found real, repeatable issues: dialogue audio degrading into repeated syllables, inconsistent sports/object physics, occasional extra fingers or malformed objects, and drifting on-screen text. These are model-level behaviors, not artifacts of running locally — we have not independently verified whether the hosted 2K regeneration pass changes this pattern, so don't assume paying for the online version fixes reliability issues by itself. Budget for reviewing every clip either way.
Decision Matrix by Use Case
| Your situation | Recommended path |
|---|---|
| One-off video, no GPU, want it fast | Online |
| Need 2K output specifically | Online — local alone won't get you there |
| High volume, already own a 24GB+ GPU | Local (ComfyUI setup) |
| 8–16GB GPU only | Local via GGUF, with a real quality tradeoff |
| Confidential/unreleased material, need it to never leave your network | Local |
| Based in the US, EU, UK, or South Korea | Online — the open weights aren't licensed for local use in these regions; see License |
| Want to fine-tune, integrate into a custom pipeline, or batch-automate | Local |
| Not technical, don't want to manage GPUs/quantization | Online |
A Hybrid Approach
Some workflows benefit from both: prototype cheaply and privately at 768p locally to lock in composition, camera direction, and pacing, then run the final selected shot through the hosted 2K version for delivery quality. This avoids paying credits on early drafts while still getting 2K for what actually ships — worth considering if you iterate heavily before committing to a final clip.
FAQ
Is running MiniMax H3 locally actually free?
No license fee, and electricity cost is negligible — but you need a capable GPU (24GB-class for the tested full-precision path, or a lower tier with GGUF and a real quality tradeoff) and your own time to set up and QA every output.
Is the online version worth it if I already have a good GPU?
It depends on your volume and whether you need 2K. If you're generating occasionally or need 2K specifically, online is simpler. If you're running high volume and 768p is acceptable, local is cheaper once set up.
Does the online version fix the audio/quality issues found in local testing?
We haven't independently verified this — our tested data is from the local RTX 4090 path. Treat every clip from either path as needing review until proven otherwise for your specific use case.
Can I use MiniMax H3 locally if I'm in the US, EU, UK, or South Korea?
Not under the Community License covering the open weights — those regions are excluded. The hosted online version operates under separate terms and is the practical option if you're in one of these regions. Full detail: License & Restrictions.
What does "minimax h3 comfyui pricing" even mean — isn't ComfyUI free?
ComfyUI itself is free and open-source. The real cost of the local path is the GPU hardware and your time, not ComfyUI. If you're comparing that against buying credits for the hosted version, see the cost breakdown above.
Ready to Try the Hosted Version?
Skip the GPU, quantization decisions, and manual QA — generate with MiniMax H3 online at up to 2K, starting with free credits.