What MiniMax released
On 11 November MiniMax published H3-Image, an image generation and editing model built on the company's H3 video model (MiniMax blog). The weights are on Hugging Face. MiniMax also sells the model through its API Platform and uses it inside Hailuo AI and MiniMax Code. The headline claims are native 4K output in under ten seconds and prices at 50 to 70 per cent of comparable mainstream services. The blog also claims "state-of-the-art prompt adherence among open-weight models" but doesn't name a benchmark, so we're treating that line as marketing until someone tests it independently.
How the model works
H3-Image is a diffusion transformer (DiT). A DiT starts from random noise and removes some of the noise at each step, with a transformer deciding what to remove. The model card lists 33.1 billion parameters across 50 layers. The Qwen3-VL text encoder comes over unchanged from H3, and the audio layers have been removed. The image autoencoder turns each 32 by 32 pixel patch into one token, so a 4096 by 4096 image comes to 16,384 tokens (Neurohive).
MiniMax distilled the model down to six sampling steps and dropped classifier-free guidance (CFG). CFG runs the model a second time at each step to push the output closer to the prompt. Fewer steps and no second pass mean less compute per image, and the cost shows up in what you can control. The guidance scale is fixed, and the card says any value other than 1 returns an HTTP 400 error. Negative prompts aren't supported either. The model works best near 16 megapixels, within a working range of about 4 to 20 megapixels, and cuts prompts off at 4,096 tokens. The card says the model expects a detailed English description of the whole image. For short or non-English prompts, it recommends running them through MiniMax's prompt expander first, which is available on the API or as a local model derived from Qwen3.8-27B.
Speed and quality claims need checking
The blog says a 4K image takes under ten seconds. Neurohive reports MiniMax's own median (P50) latency chart as 16.0 seconds in Standard mode and 10.3 seconds in Highspeed mode at 4K, so the sub-ten-second figure looks like a best case. Neurohive also points out that MiniMax re-measured H3-Image in a second round while the competitors' times came from the original comparison, which makes the speed gap less reliable than the chart suggests.
MiniMax hasn't published a technical report, a description of its training data or any quality metrics, and the model wasn't in the Artificial Analysis rankings as of 11 November (Neurohive). Early hands-on reports collected by Neurohive are mixed. Testers liked the natural-looking portraits and the character edits from a single reference image. Small text came out garbled on infographics, and in one product mockup H3-Image dropped details that Seedream 5.5 kept. Most of these reviewers had early access, and one post was labelled as a paid partnership.
API access, pricing and limits
The API has two models, MiniMax-H3-Image and MiniMax-H3-Image-Highspeed, both served on /v1/images/generations and /v1/images/edits (MiniMax image generation guide). A request can set an area of 1K, 2K or 4K, with 4K as the default. It can also set a custom size between 655,360 and 20 million pixels, with an aspect ratio from 1:3 to 3:1, and ask for up to ten images at once. Edits accept one to 16 reference images, and the prompt refers to them by upload order. The guide recommends a timeout of at least 600 seconds and turning off SDK retries because generation isn't idempotent. A batch request returns HTTP 200 when at least one image succeeds, so client code needs to check each item for an error field.
Standard text-to-image prices per image on the pricing page:
- 1K: $0.02179 (Highspeed $0.03529)
- 2K: $0.02333 (Highspeed $0.03622)
- 4K: $0.03439 during the launch discount, normally $0.06879 (Highspeed $0.06501, normally $0.13001)
The 50 per cent discount applies only to outputs of at least 16.6 million pixels. The discount ends at 16:00 UTC on 17 November, which is 5am on Wednesday 18 November in New Zealand. After that, a 4K Standard image costs about three times as much as a 2K image. Edits are billed per token at the following rates:
- Input text: $2.30 per million tokens on both tiers
- Input images: $1.50 per million tokens on Standard, $3.00 on Highspeed
- Output images: $3.80 per million tokens on Standard, $7.60 on Highspeed
MiniMax's own example prices a 4K edit with three reference images at $0.07914 on Standard at full price. At 1K, Highspeed costs about 1.6 times as much as Standard. MiniMax's chart puts Highspeed's time saving at 4K at about 5.7 seconds per image.
Self-hosting and the licence terms
Running the model yourself means downloading a 66.3 GB DiT plus a 66.7 GB text encoder from the main H3 repository. The model card shows single-GPU serving with SGLang:
sglang serve --model-path MiniMaxAI/MiniMax-H3-Image --host 0.0.0.0 --port 30000 --warmup-mode offThe card names NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X as supported hardware. Neurohive reports two community runs on cheaper machines. An RTX 3090 with INT8 weights produced a 4K image in 55.79 seconds, and a 64 GB Mac mini M5 Pro took 547 seconds.
The weights come under the MiniMax-H3-Image Community License, which reads very differently from Apache 2.0 (licence). The licence covers every territory except the EU, the UK, South Korea and the US. Use, distribution or display in those regions isn't authorised. You need prior written permission from MiniMax if your annualised revenue, including affiliates, is above US$1 million, or if you run a hosted service for third parties. Commercial products must show the name "MiniMax-H3-Image" prominently in their interface. You can't use the model or its outputs to improve any other AI model, and the licence counts distillation and training on synthetic data as derivatives. Hong Kong law governs the agreement. MiniMax claims no rights over the images you generate.
What this means for New Zealand builders
New Zealand is inside the licence territory, so a local team can test the weights without asking anyone. A product that shows generated images to customers in the US or UK falls outside the self-hosting licence, though, and so does any hosted image service for clients unless MiniMax gives permission. Those rules apply to the weights. API use falls under MiniMax's platform terms, which this article doesn't cover, so read those terms before you ship anything that calls the API.
For most teams the API is the practical route. A self-hosted setup needs more than 130 GB of downloads and a data-centre GPU to reach the advertised speeds. For anyone already generating large images for print, retail catalogues or game assets, $0.069 for a 4K Standard image at full price is cheap enough to compare against current suppliers. Test it on the jobs where testers found weaknesses: small text, keeping product details intact, and keeping an illustration style consistent across poses. The training restriction rules out using H3-Image outputs as synthetic data to train another model, so it won't work as a data generator for fine-tuning projects. If you put H3-Image through its paces on a local project, you can share the results through ASAI projects.
Sources
- MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era (MiniMax blog)
- MiniMaxAI/MiniMax-H3-Image model card (Hugging Face)
- MiniMax-H3-Image Community License (Hugging Face)
- Image generation guide (MiniMax API Platform docs)
- Pricing overview (MiniMax API Platform docs)
- MiniMax H3-Image: Open-Weight Model Generates Native 4K Images in 10-16 Seconds (Neurohive)
- LLM Gateway model timeline
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.