MiniMax H3-Image: Price, API, Test Results and Comparison
MiniMax H3-Image in one place: what it does, where to use it, what it costs, the licence and my test of 170+ images against Seedream 5.5, Nano Banana 2.1 and GPT-Image-2.6.
👁 152 reads
MiniMax H3-Image is an image generation and editing model from MiniMax, the company behind Hailuo AI and the H3 video model. It launched on 10 November 2026, outputs about 16 megapixels natively, accepts up to 16 reference images and has open weights on Hugging Face. On launch day I ran more than 170 images through its API for 4.98 USD. This guide collects the facts, the prices and my results in one place.
The short version: a very good model for product photos, illustrations, covers and scenes built from many references, and cheap in 4K during launch week. It is not the right tool for text with diacritics (Nano Banana 2.1 and GPT-Image-2.6 do better), it cannot make transparent backgrounds, and its weights licence excludes several regions.
What MiniMax H3-Image is
According to the Hugging Face model card:
- the diffusion transformer has 33.1 billion parameters (50 layers); the text encoder is the Qwen3-VL encoder taken unchanged from MiniMax H3;
- generation and editing share one checkpoint; editing accepts 1 to 16 images;
- about 16 MP in any aspect ratio, from 4096×4096 to 6272×2688, without an upscaler; the working range is 4-20 MP;
- distilled to 6 steps, no CFG and no negative prompt;
- there is a Highspeed version and an optional prompt expander (MiniMax-H3-Image-PE, 17.9 GB) that turns a short prompt in any language into a long English description.
Where it works and what it costs (10 Nov 2026, evening)
| Provider | 1K / 2K | 4K (16 MP) | Editing | Notes |
|---|---|---|---|---|
| MiniMax API (platform.minimax.io) | 0.02179 / 0.02333 USD | 0.06879 USD, with the offer 0.03439 USD | 1 reference in 4K 0.03656 USD with the offer | token pricing; offer until 17 Nov 2026 16:00 UTC for images of at least 16.6 million pixels |
| MiniMax API Highspeed | 0.03529 / 0.03622 USD | by tokens (output image 7.60 USD/M) | yes | not faster at 2K in my test |
| fal.ai | same token rates as MiniMax | yes | Standard and Highspeed since 10 Nov | |
| WaveSpeedAI | 0.03 USD | 0.09 USD | from 0.038 USD | flat price, no launch discount |
| Hailuo AI (app) | 3 credits per image in 1K, 2K and 4K | up to 16 images | free starter credits, prompts up to 7,500 characters | |
| OpenRouter, Replicate | not available | - | checked on the evening of 10 Nov | |
Watch the offer threshold: my 4:5 portraits at 3648×4544 (16.58 MP) were charged the full 0.067-0.070 USD, while 5472×3072 and 4096×4096 cost 0.033-0.036 USD.
My test in numbers
| Test | MiniMax H3-Image | Seedream 5.5 | Nano Banana 2.1 | GPT-Image-2.6 |
|---|---|---|---|---|
| Polish text, 7 prompts, errors per run (about 550 characters) | 26 (25, 24, 29), 0 of 21 images clean | 8 (6-10) | 0 (5 prompts, 2 runs) | 1 in 2 runs (5 prompts) |
| 10 short slogans x 3 | 22 errors in 453 characters, 17/30 clean | 6 in 453, 27/30 | 0, 30/30 | 0 in 151, 10/10 |
| Transparent background | 0/5 (not in the API) | 10/10 RGBA | - | - |
| Face from 3 photos, ArcFace (10 images) | 0.573 | 0.632 (5 images) | 0.508 | 0.729 |
| 8 chained edit rounds | all edits done, texture builds up from round 2-3 | 8 rounds without artefacts | - | - |
| Time 2K / 4K (API response) | 13-17 s / 22-24 s | about 21-28 s (2K) | 13-19 s | 24-43 s |
Competitor results come from my own tests on 9-10 November with exactly the same prompts. The face test uses photos of a person who agreed to have his likeness used in my tutorials; ArcFace is computed the same way for every model.
What it is good for, and what not
- Yes: product and e-commerce photos, covers and posters with English text, storyboards, illustrations in many styles, interiors composed from many items, large-format print (16 MP straight away).
- Rather not: graphics with text that needs diacritics, cut-outs on transparent backgrounds, long step-by-step editing of one image, a faithful likeness of a specific person (GPT-Image-2.6 holds it better).
Open weights and the licence
The weights are in the MiniMaxAI/MiniMax-H3-Image repository, and on day one the community already published ComfyUI versions (int8 21 GB, w4a8 12.6 GB) and GGUF files (from 6.9 GB). The MiniMax-H3-Image Community License, however, excludes the European Union, the United Kingdom, South Korea and the USA from the territory where the weights and their outputs may be used, and companies with more than 1 million USD in annual revenue or hosted services need separate permission. Readers elsewhere, for example in India, Canada or Australia, may run the weights under the licence terms. I work from Poland, inside the EU, so I tested the model only through the API.
More from my test
The launch-day news with the first results is here: MiniMax H3-Image released. All 24 gallery prompts with images, time and cost are here: 24 tested MiniMax H3-Image prompts. Prices at every provider and the 16.6 MP trap: MiniMax H3-Image price and API. The full four-model test: MiniMax H3-Image vs Seedream 5.5, Nano Banana 2.1 and GPT-Image-2.6.
Frequently asked questions
Is MiniMax H3-Image free?
Hailuo AI gives new users free credits, and one H3-Image picture costs 3 credits at any resolution. Through the API you pay from the first image: about 0.023 USD for 2K and 0.034 USD for 4K until 17 November.
Can it write text in my language?
You can prompt in any language because the prompt expander rewrites it into English. Text inside the image is another story: English was clean in my tests, while Polish with diacritics averaged 26 errors per 550 characters.
Does it support transparent backgrounds?
No. The API documentation lists transparent backgrounds as unsupported, and 5 out of 5 of my images had a full background.
How long does a 4K image take?
22-24 seconds via API, 2K 13-17 seconds. MiniMax quotes about 10 seconds, probably on its own infrastructure without a queue.
How is it different from MiniMax H3?
MiniMax H3 is a video model with sound. H3-Image uses the same architecture without the audio branch and only makes images.
Sources
- Hugging Face: MiniMaxAI/MiniMax-H3-Image (model card, 10 Nov 2026)
- MiniMax-H3-Image Community License Agreement
- MiniMax: Image Generation Guide
- MiniMax: API pricing and the 4K offer (10 Nov 2026)
- WaveSpeedAI: MiniMax H3 Image (10 Nov 2026)
- fal.ai: MiniMax H3 Image (10 Nov 2026)
- Comfy-Org: ComfyUI files
- Hailuo AI
- My own test on 10 Nov 2026 via API: more than 170 images, cost from the API cost field (4.98 USD), text read at full zoom, ArcFace (insightface buffalo_l). Competitors: my tests on 9-10 Nov 2026 with the same prompts.
More on this from QUANTSLANT
news·MiniMax H3-Image Is Out: 16 MP, Open Weights, Price and My TestMiniMax released H3-Image: 16 MP images, up to 16 references, open weights and 4K for 0.034 USD in launch week. I ran 170+ images through it.