MiniMax H3 Image vs Its Siblings: Capability, Price and Which Task Each One Fits

Short answer

H3 Image (h3-image) is, as of this launch, the only MiniMax model whose job is producing pixels. So "which one" is really two questions: does this step need an image at all, and if so, which route gets you one.

• You need an actual image → H3 Image. Native 4096x4096, roughly 16 megapixels, under 10 seconds per image by MiniMax's own figure, up to 16 reference images, weights published.
• You need prompts, a numbering scheme, a generation pipeline → minimax-m3.1-flash. Already on TeamoRouter, same key, but it draws nothing.
• You need heavier text reasoning or gnarlier prompt engineering → MiniMax-M3. The bigger model in the same family. Also text-only.

h3-image is not on TeamoRouter yet. Once it is live, the same key and model: h3-image is all you change; until then, use minimax-m3.1-flash. For a 4K image today you have two options: Hailuo AI, or the weights on your own GPU.

What each model is for

H3 Image: the pixel step

Its three claims only work as a set. 4K decides whether detail survives a zoom, 16 references decide whether you can specify "these people, this palette, this arrangement", and 10 seconds decides whether rerolling five times for one detail is tolerable.

The tasks that keep showing up in MiniMax's and the community's demos:

• Map, position, heading → the matching view. Mark a point, give a direction, get what you would see standing there. For fiction scenes and game storyboards this beats repeating "a bit more left, a bit further".
• Materials and lighting onto a blockout. Keep the composition, resolve metal, glass, rubber and ground dust separately, with light and shadow.
• Fabric, skin and makeup detail. Stitching, wear and creases at a collar, skin texture, cracked makeup. Naming those objects beats writing "cinematic".
• Group shots with assigned palettes. One reference collage with eight people and eight colour cards, matched by index, each getting a primary, secondary and accent colour while stance and spacing vary too.
• Multi-view character sheets. Front, back and both profiles as black line art on white, aligned at equal height.
• Furniture in an empty room. A room plus a few pieces, positions spelled out, so you can see them together before ordering.
• Chinese layouts with the reasoning included. Conditions, candidates and the worked logic in one image.

minimax-m3.1-flash: clear the work before the pixels

It cannot take over the image step, but it can take everything around it: turn eight people and eight colour cards into an indexed mapping, spell out who is number what and which palette they get, break a brief into per-object prompt fragments, and lay out a batch job list. Do that prep work first and you spend far fewer generations later.

MiniMax-M3: when the text side gets hard

The heavier tier in the same MiniMax family. Reach for it when prompt structure is complex, when one setting has to be split into many characters and palettes, or when you are writing the script that wires generation into an agent. Still text-only.

Pricing: the word "paid", and no numbers

This section needs to be blunt, because "per-image price" claims are already circulating.

• The launch material contains no API pricing. No per-image rate, no free allowance, no concurrency or rate limits.
• Cloud use goes through Hailuo AI, which MiniMax explicitly calls a paid feature. No unit price, no free tier stated.
• Running the weights locally is not billed per image. It costs a GPU and time, and there is a ready ComfyUI template.

One distinction to keep straight: a paid tier inside Hailuo AI is not the same thing as per-image API billing. You cannot infer one from the other. Until MiniMax publishes API pricing, no specific per-image figure has an official basis, and we are not inventing one.

Three ways to get an H3 Image

Same model, three very different cost structures:

The local route is gated by hardware: ComfyUI 0.45.0 or newer, template package 0.11.104 or newer. If the official weights will not fit your VRAM and RAM, find a quantised build for your card instead of forcing it. And do not quote MiniMax's 10 seconds as your own number — whatever your GPU does is your real speed.

A practical order of decisions

Do not start by agonising over model choice. Work down this list:

1. Does this step need pixels? No — prompts, numbering, batch splitting, agent wiring — go straight to minimax-m3.1-flash, and MiniMax-M3 if it gets complicated.
2. Pixels, and you have a GPU? Run it locally. Weights are out, no per-image billing, and rerolling is cheap.
3. Pixels, no GPU or no appetite for setup? Hailuo AI. MiniMax says it is paid; the price is whatever they publish.
4. Neither appeals? Wait for h3-image on TeamoRouter. Meanwhile have minimax-m3.1-flash write the prompts and the numbering scheme so you can pick straight up.
5. Heavy Chinese text in the image? It lays out Chinese, but text-dense scenes drop characters. Budget rerolls.

Using them on TeamoRouter

h3-image is not in our catalog yet. The two text models are callable today:

from openai import OpenAI

client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
    model="minimax-m3.1-flash",  # switch to MiniMax-M3 for heavier reasoning
    messages=[{
        "role": "user",
        "content": "Turn the eight people and eight colour cards in this reference collage into an indexed mapping, then write an outfit prompt per person with primary, secondary and accent colours spelled out",
    }],
)
print(resp.choices[0].message.content)

h3-image is not on TeamoRouter yet. Once it is live, the same key and model: h3-image is all you change; until then, use minimax-m3.1-flash. The base URL stays put — it runs through our standard image generation entry point, not a separate auth path.

How to pay

No subscription: top up your balance in the console from $5 and each request is billed against it by actual usage. Cards are supported, and so are USDT, USDC and ETH. The same key also calls Claude and GPT models.

New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.

FAQ

H3 Image or minimax-m3.1-flash — which do I use?

Ask whether the step produces an image. If it does, only H3 Image will do. If it is prompts, numbering or scripts, minimax-m3.1-flash is faster and cheaper.

Can I call all three on TeamoRouter right now?

No. minimax-m3.1-flash and MiniMax-M3 are available; h3-image is not. It shipped, but it has not entered our catalog. Once it does, the same key and a new model ID is all it takes.

How much does H3 Image cost?

MiniMax has not published API pricing. Confirmed: local inference is not billed per image, and cloud use through Hailuo AI is a paid feature with no stated rate or free allowance.

Is Hailuo AI's paid tier the same as the API price?

No. Hailuo AI is the cloud product MiniMax names, and they only say it is paid. Per-image API billing has no published figure at all. You cannot infer one from the other.

I want a 4K image today. What are my options?

Two: deploy the weights locally (no account, no card, needs a GPU), or use Hailuo AI on its paid plan. Hand the text side to minimax-m3.1-flash or MiniMax-M3 meanwhile.

What does local deployment require?

ComfyUI 0.45.0 or newer and template package 0.11.104 or newer. If VRAM or RAM is tight, use a quantised build suited to your card.

Is Chinese text in the image reliable?

Not really. It lays out Chinese, but text-heavy scenes lose a few characters and need rerolls. Build a few extra attempts into the budget for finished graphics.

Related reading

• MiniMax H3 Image is out: model ID, price and how to call it: specs, the unpublished-pricing situation and both call formats
• Using MiniMax H3 Image from China and Russia: what you can do without an overseas account or a credit card