AI News · 5 min read
MiniMax H3-Image: open-weight 16MP image model, with a catch
MiniMax released H3-Image on November 10 with open weights, 16-megapixel output and up to 16 reference images. Prices, hardware needs and the license limits.
What MiniMax released
On November 10, 2026, MiniMax launched MiniMax H3-Image, an image generation and editing model built on its H3 video model. "4K image generation in 10 seconds," the company posted on X ↗. "And it's open-weight." China's Yicai ↗ reported the release the same evening, and the weights are up on Hugging Face ↗.
You can use it three ways: through the MiniMax API, in the Hailuo AI app, or by downloading the weights and running it yourself.
Here's what the model card ↗ says:
About 16 megapixels at every aspect ratio, from 4096×4096 (1:1) to 6272×2688 (21:9), "directly and without an upscaler." MiniMax says it works best around 16MP and works from about 4MP to 20MP.
One checkpoint for both jobs. It handles text-to-image and editing with up to 16 reference images.
Six sampling steps. The checkpoint is distilled to 6 steps with no classifier-free guidance, which is how MiniMax claims its speed.
A 33.1B-parameter diffusion transformer. It reuses the text encoder from MiniMax-H3 (based on Qwen3-VL) without changes.
The speed and quality claims are MiniMax's own. We haven't seen independent benchmarks yet.
What it costs on the API
MiniMax's pricing page ↗ bills by tokens, then gives example prices per image. Two models are listed: standard MiniMax-H3-Image and a faster MiniMax-H3-Image-Highspeed at twice the image-token price.
| Job (MiniMax's examples) | Standard | Highspeed |
|---|---|---|
| Text-to-image, 1K (1024×1024) | $0.02179 | $0.03529 |
| Text-to-image, 2K (2048×2048) | $0.02333 | $0.03622 |
| Text-to-image, 4K (4096×4096) | $0.06879 ($0.03439 at launch) | $0.13001 ($0.06501 at launch) |
The launch discount takes 50% off any image of at least 16.6 million pixels until November 17, 2026 at 16:00 UTC (21:00 PKT). Per million tokens, Standard costs $2.30 for text input, $1.50 for image input and $3.80 for image output. Highspeed is $2.30, $3.00 and $7.60. One detail from the API guide ↗: requests aren't idempotent. If your client times out, MiniMax still finishes and bills the image, so turn off automatic retries.
Running it yourself is a big download
"Open weights" doesn't mean it will run on a gaming PC. The model card lists the diffusion transformer at 66.28 GB in BF16. The required text encoder is another 66.7 GB, downloaded from the MiniMax-H3 repo. An optional prompt expander, MiniMax-H3-Image-PE, adds 17.9 GB.
The model card says it works with SGLang, vLLM-Omni, diffusers and ComfyUI. The vLLM recipe names NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X among supported accelerators. That's datacenter and workstation hardware. MiniMax also says the model expects "a detailed English description of the whole image." Short prompts go through a prompt-enhancement step on its API first.
The license is the real story
The weights come under the MiniMax-H3-Image Community License, not Apache or MIT. We read the license text ↗. Four clauses matter most:
Territory. The license only applies outside its "Excluded Territories," which it defines as "the European Union, the United Kingdom, the Republic of Korea and the United States of America." Use in those places "is not authorized by this Agreement." MiniMax invites people there to contact it for a separate license.
Revenue and hosting. You need written permission from MiniMax if your company's annualized revenue is over $1 million, or if you offer the model as a hosted service to others.
Attribution. A commercial product that uses it has to show "MiniMax-H3-Image" in its interface.
No training other models. You can't use the weights or their outputs to improve any other AI model.
So most readers of this site can't legally self-host it under this license. That covers anyone in the US, UK or EU. MiniMax's own API and Hailuo app are separate hosted services with their own terms.
What people are saying
Early X reaction was loud, and much of it was creators posting sample images. A few posts were about more than the pictures:
The vLLM team announced day-0 support ↗ in vLLM-Omni "on both NVIDIA and AMD GPUs," and VeRL-omni added LoRA training support ↗ on the first day.
Runware put both versions live ↗ on launch day. It noted that MiniMax calls H3-Image "the strongest open-weight image model released so far," which is MiniMax's claim, not a test result.
Developer CoderGuLu summed up the fine print ↗: "$1M+ annualized revenue or a hosted service needs written authorization. US, EU, UK, and Korea are excluded."
ChinaHotTech asked the practical question ↗: now that the weights are out, it's "whether that checkpoint actually runs locally at a usable VRAM budget."
We couldn't reach Reddit for this story, so these reactions come from X only.
Our take
If you want cheap 4K images by API, H3-Image is worth trying this week. MiniMax lists about 3.4 cents for a 4096×4096 image during the launch discount and about 6.9 cents after it. For poster, print or product work that needs native 16MP without an upscaler, that's a reason to test it against what you use now. Reference-heavy jobs like product shots, character sheets and room layouts are where up to 16 references could matter most. Test it on your own prompts before you switch.
If you wanted a new local model, look elsewhere. About 135 GB of required downloads (the 68.65 GB repo plus the 66.7 GB text encoder) and a license that excludes the US, UK and EU make this a datacenter release for most people. For image models you can run at home, start with our Stable Diffusion tutorial and ComfyUI guide.
Closed alternatives: Our Midjourney alternatives roundup covers the main hosted generators. ByteDance's Seedream 5.5 is the closest rival for multi-round editing, and OpenAI's ChatGPT Images 2.6 is the easy choice if you already pay for ChatGPT. You'll find more tools under AI image generation.