🗺️ Where to Run MiniMax H3: Every Route Compared

H3 shipped with more places to run it on day one than most models get in a year. That is unusual, and it means the question stopped being "can I get access" and became "which route, and what does each one cost me in flexibility."

By the numbers

Route	What it is	Notable limits
MiniMax Open Platform API	The source. v2 endpoints, full spec	Region-split hosts; base pricing
Hailuo AI app	Consumer app, no code	Credit-based, app-side features
MiniMax Hub	Desktop app	Consumer-oriented
fal	Official API partner	Three separate endpoints by mode
ComfyUI Partner Nodes	Node graph integration	Hosted API call, not local
Vercel AI Gateway	Drop into a Next.js app	Gateway routing
OpenRouter	Unified API, slug minimax/hailuo-3	Single upstream provider
Atlas Cloud / EvoLink / Apiframe / kie.ai	Resellers	Their margin, their base URL

Go direct if you need the full spec

MiniMax's own v2 API is the only route that documents the complete capability surface, and it is the one we built our tooling against.

That surface: model ID MiniMax-H3, 2K as the only accepted resolution, 24fps fixed, 4–15 second integer durations, and a prompt budget of 7,000 characters. Five input modes including last-frame-only, which most gateways do not expose. Reference inputs capped at 9 images, 3 videos and 3 audio clips with a hard 12-file total.

The setup detail that catches people: MiniMax runs region-split hosts. api.minimax.io is global, api.minimaxi.com is mainland China, and a key from one against the other returns 401 authorized_error. That single mismatch accounts for most "invalid API key" reports. The Python setup guide covers the rest.

fal, ComfyUI and Vercel: the integration routes

fal launched as an official API partner with three dedicated endpoints — text-to-video, image-to-video, and reference-to-video. Splitting by mode is a design choice worth understanding: it makes each endpoint's schema simpler, and it means you pick the mode by URL rather than by what you put in the request body.

ComfyUI shipped Partner Node support at launch, documenting multimodal I/O, native stereo audio on every clip, 2K at 5–15 seconds and 24fps, and in-place editing. Read their wording carefully though — the announcement says API access is live today and native open-weight support is coming. Those are different things, and the distinction is the single most misunderstood point about H3 right now.

Vercel AI Gateway carries H3 for text, image or reference-driven generation. If your product is already a Next.js app on Vercel, this is the shortest path from nothing to a generated clip.

OpenRouter lists it as minimax/hailuo-3 — useful if you are already routing other models through it, and cheap to A/B against Veo 3.1's published rates. It notes the requests forward to a single upstream provider — so you are getting MiniMax's own service with OpenRouter's billing and key management on top.

The reseller tier, and how to read its pricing

Atlas Cloud, EvoLink, Apiframe and kie.ai all carry H3 behind their own API keys and base URLs. They are legitimate routes, particularly if you are already consolidating spend through one of them. They are also where most of the conflicting numbers on the internet come from.

Observed price quotes for the same 2K tier range from $0.13 to $0.14 per second, with 768p quoted at $0.09 or $0.10. Apiframe bills in credits — 22 credits per second of 2K output — which makes direct comparison harder rather than easier.

MiniMax has published no official H3 rate card. Until it does, treat every number you see as a reseller's rate including their margin, and read the real figure off usage.total_seconds on each finished task. Remember that reference-video seconds bill as input on top of your output seconds.

The local question, answered plainly

You cannot run H3 on your own hardware today.

MiniMax's launch framed the model around open weights and stated an intent to publish them within days under its Community License. SCMP and Dealroom both reported the plan. Nothing has shipped. When it does, the two places it would surface are MiniMaxAI's Hugging Face org and native ComfyUI support.

Until then, "ComfyUI supports H3" means a node that calls a paid hosted API. It will not work offline, it consumes credits, and it does not care what GPU you own. For what genuinely runs locally today, see open-weight AI video models.

Pick your route

• Cheapest per clip → MiniMax direct. Everything else is a margin on this.
• Fewest integration steps in an existing app → Vercel AI Gateway if you are on Next.js, OpenRouter otherwise.
• Already living in a node graph → ComfyUI Partner Nodes.
• Want mode-specific endpoints and good docs → fal.
• Need last-frame-only or the full 12-file reference stack → MiniMax direct; gateways vary. The omni-reference workflow explains how to spend that budget.
• Waiting to self-host → nothing to do yet. Watch Hugging Face.

How we checked this

Route list and capability claims come from each platform's own launch material: fal's partner announcement and model pages, ComfyUI's Partner Node post, Vercel's changelog entry, and OpenRouter's model listing. API limits are read off MiniMax's v2 reference, which we reconciled endpoint by endpoint while building our own client.

Pricing is the weak spot and we have said so rather than picking a number. The conflicting reseller quotes are reported as conflicting.

Last verified: August 1, 2026. This page has a short shelf life. Three things would change it: the weights actually dropping, 768p leaving closed beta, and more hosts adding the model. We are not stating a resolution or price we have not seen published.

What we left out

Per-platform latency figures. Nobody has run enough volume on a two-day-old model to publish render times that mean anything, and repeating a launch-week number as a benchmark would be guessing.

We also left out a "best platform" verdict. The right route depends on where your code already lives, and a page that picks a winner here is usually picking the one paying it.