Short answer
MiniMax released H3 Image on November 11, 2026. The model ID is h3-image. It is an open-weights text-to-image and image-editing model built around three claims: native 4K output, under 10 seconds per image, and up to 16 reference images.
The facts, all from the launch material:
- Output: native 4096x4096, roughly 16 megapixels. MiniMax's own figure is one 4K image in under 10 seconds.
- Reference images: up to 16 in a single pass. People, outfits and colour palettes can be supplied separately and tied together by number.
- Weights: published. You can run it locally, and ComfyUI ships a ready-made template.
- Cloud: available in Hailuo AI, which MiniMax describes as a paid feature.
- API pricing: not published. No per-image rate, free allowance or rate limit has been given.
h3-image is not on TeamoRouter yet. It shipped, but it has not reached our catalog. Once it does, the key you already have works: set model to h3-image and keep the base URL. In the meantime, on the text and agent side you can use minimax-m3.1-flash or MiniMax-M3 — those are language models and will not draw anything for you. To generate images today, you need to deploy the weights locally (ComfyUI) or use Hailuo AI.
What was actually announced
H3 Image is not another "it can draw" model. MiniMax positions it as one that survives a real production pipeline: high enough resolution to inspect, enough reference images to explain a brief, and fast enough that rerunning it five times chasing a detail is not painful. The three only matter together — 4K decides whether detail holds up when zoomed, 16 references decide whether you can specify "these people, this palette, this arrangement", and 10 seconds decides whether you are willing to reroll.
The other headline is that the weights are open. Paying per generation in the cloud is right for trialling and light use. Once you care about cost, offline operation, or wiring generation into your own pipeline, running the weights yourself is the normal answer. H3 Image leaves both doors open.
What the official demos show
The examples MiniMax and the community published map closely onto the work this class of model actually gets used for:
- Map, position and heading into a matching scene. Drop a marker on a map, give a direction, and ask for what you would see standing there facing that way. For novel scenes and game storyboards this beats repeating "a bit more left, a bit further away".
- Materials and lighting onto a grey box. Hand it a blockout with the composition already set, keep the layout, and ask it to resolve metal, glass, rubber, ground dust and light and shadow separately. Common for sci-fi environments and product concepts.
- Fabric and skin detail. Stitching, wear and creases at a collar, skin texture, cracked makeup. Naming those objects in the prompt does more than writing "cinematic" and hoping.
- Group shots with assigned palettes. One reference collage holding eight people and eight colour cards, matched by index, so each person gets a specified primary, secondary and accent colour while stance, seating and spacing get varied too. Work that used to be one image at a time now lands in one pass.
- Multi-view character sheets. Upload a character, ask for front, back and both profiles as black line art on white, arranged at equal height.
- Furniture in an empty room. Give it a room and several pieces, state where each goes, and see whether they get along before anything is delivered.
- Chinese layouts with the reasoning included. One official demo is a staff-selection logic puzzle: Chinese conditions, the selected and rejected candidates, and the worked reasoning, all laid out in one image.
Limits worth knowing
- Chinese text still drops characters. MiniMax admits this is a long-standing weak spot: text-heavy scenes occasionally get a few characters wrong and need a reroll. Do not plan on a one-shot finished graphic with Chinese copy.
- 16 reference images is a hard ceiling. Beyond that you need to compose your own collage and merge like references.
- Local deployment has version requirements. For ComfyUI you need 0.45.0 or newer, and the template package needs 0.11.104 or newer. If the official weights will not fit your VRAM and RAM, find a quantised build suited to your card rather than forcing it.
- Speed depends on hardware. The 10-second figure is MiniMax's. Local throughput varies enormously with the GPU, and community numbers diverge from vendor numbers by multiples as a rule.
Pricing
Be precise about this: the launch material contains no API pricing.
Two things are stated:
- Local use is not billed per image. The weights are out; self-hosting costs you a GPU and time, and the ComfyUI template is provided.
- Cloud use through Hailuo AI is paid. MiniMax says only that it costs money — no per-image rate, no free allowance.
Per-image price, whether a free tier exists, and concurrency or rate limits are all unpublished. We are not guessing any of those numbers. When MiniMax publishes pricing, or when h3-image is available on TeamoRouter, this section gets the real figures and the billing model.
Calling it on TeamoRouter (once it lands)
Not available yet. When h3-image enters our catalog, your key and base URL stay put — only model changes.
OpenAI-compatible format, using the standard SDK:
from openai import OpenAI
client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.images.generate(
model="h3-image",
prompt="View from the red arrow on the map, looking the way it points: suspension bridge over a river, waterfall and a castle on the hill, game render style",
size="4096x4096",
)
print(resp.data[0].url)
The vendor-native format hits the same host, using MiniMax's image_generation endpoint:
curl https://api.teamorouter.com/v1/image_generation \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "h3-image",
"prompt": "Resolve materials and lighting on this blockout; treat metal, glass and ground dust separately"
}'
The two are different entry points only. Same model, same key, same billing. Exact parameter names and response shapes will follow our docs at launch.
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.
What to use today
minimax-m3.1-flash: available on TeamoRouter, fast and cheap — good for getting your prompts right and wiring the generation flow into an agent. It does not produce images, but it can write the prompts, the numbering scheme and the batch jobs for you.MiniMax-M3: the heavier model in the same MiniMax family. Reach for it when you need stronger reasoning or more elaborate prompt engineering. Also text-only.
There is no stand-in for the image step itself. To produce 4K images immediately you have two options: deploy the weights locally, or use Hailuo AI on its paid plan.
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
Is MiniMax H3 Image out?
Yes. It launched on November 11, 2026 under the model ID h3-image, with the weights published at the same time.
Is H3 Image open source?
Yes. MiniMax released the weights, so it can be deployed locally, and there is a ready-made ComfyUI template (ComfyUI 0.45.0+, template package 0.11.104+).
How much does H3 Image cost?
MiniMax has not published API pricing. What is 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. We will fill in real numbers here once they are published or once the model is on TeamoRouter.
What is the model ID?
h3-image.
How many reference images can it take?
Up to 16. People, outfits and colour cards can be uploaded separately and matched by image number.
Can it render Chinese text?
It can lay out Chinese, but text-heavy scenes drop a few characters now and then and need a reroll. Budget a few extra attempts for anything with Chinese copy in it.
Can I call h3-image on TeamoRouter right now?
Not yet. It shipped, but it has not entered our catalog. Once it does, the same key and a new model ID is all it takes. For text work in the meantime, use minimax-m3.1-flash or MiniMax-M3.
Once it is live, is it just a model-name swap?
Yes. It runs through our standard image generation entry point, not a separate auth path. Same key, same base URL, only model changes.
Related reading
- Using MiniMax H3 Image from China and Russia: what you can do without an overseas account or a credit card
- MiniMax H3 Image vs other image models: the head-to-head comparison