MiniMax has released MiniMax-H3-Image as open weights on Hugging Face. The official model card describes a unified text-to-image and editing checkpoint that can take up to 16 reference images, generate about 16-megapixel pictures in six sampling steps, and run locally or through MiniMax’s API, Hailuo apps and MiniMax Code.
MiniMax’s official account says the model can produce a 4K image in 10 seconds, and that API, MiniMax Code and Hailuo AI users get “50% off 16MP+ generations during launch week.” Ronny He, head of R&D at MiniMax Code, had earlier posted that the model was coming, then wrote that it was “Now Live on @MiniMax_Code.” The Hugging Face release is dated 10 November, and MiniMax also has API docs and a MiniMax Code product page.
What the release includes
The image checkpoint is a 33.1-billion-parameter diffusion transformer from the H3 family, with the audio layers removed. MiniMax says generation and editing share one set of weights. The model is distilled to six sampling steps, with no classifier-free guidance. A separate optional prompt expander, MiniMax-H3-Image-PE, is a 27-billion-parameter model fine-tuned from Qwen3.8-27B. The text encoder is not in the image repository: model_index.json loads it from MiniMax-H3.
| Component | What MiniMax published |
|---|---|
| MiniMax-H3-Image | Text-to-image and editing; BF16 DiT, 68.65 GB |
| MiniMax-H3-Image-PE | Optional text-to-image prompt expander; NVFP4, 17.9 GB |
| Text encoder | Unchanged Qwen3-VL encoder from MiniMax-H3, 66.7 GB |
| Image VAE | CNN encoder and 2D ViT decoder, 2.34 GB |
| Sampling | 6 steps, guidance scale 1; other values are rejected |
| Reference images | Up to 16 |
The model card says it generates images of about 16 megapixels at every listed aspect ratio, “directly and without an upscaler.” The working range is about 4 million to 20 million pixels, with the best results “around 16 MP.” Common sizes include 4096×4096 for 1:1 and 5472×3072 for 16:9.
| Aspect ratio | Size (width × height) |
|---|---|
| 1:1 | 4096×4096 |
| 4:3 / 3:4 | 4736×3552 / 3552×4736 |
| 3:2 / 2:3 | 5024×3360 / 3360×5024 |
| 16:9 / 9:16 | 5472×3072 / 3072×5472 |
| 21:9 | 6272×2688 |
Where it runs
The hosted API is documented at platform.minimax.io, with a China endpoint at platform.minimax.cn. Text-to-image goes to POST /v1/images/generations, and image editing to POST /v1/images/edits. There is also a Highspeed API variant, MiniMax-H3-Image-Highspeed, which the docs say takes the same parameters and limits, and an H3-Image-Context-IR endpoint that returns an enhanced prompt without generating an image. On the API, prompts can be up to 32,000 characters; with prompt enhancement switched off, only about the first 4,096 tokens are used. Local SGLang requests also truncate the prompt to 4,096 tokens. The service does not support preview images, masked edits or transparent backgrounds.
MiniMax also points people to Hailuo on the web at hailuoai.video (and hailuoai.com in China) and to a desktop app at agent.minimax.io. Invited testers had already been posting Hailuo results before the weights appeared. For local inference, the model card recommends SGLang, vLLM, diffusers and ComfyUI. A ComfyUI pull request adding native H3-Image nodes was opened the same day.
API pricing
Hosted image generation is billed in tokens, in US dollars, not at a fixed price per picture. MiniMax’s pricing overview lists these rates:
| Model | Input text, $ / M tokens | Input image, $ / M tokens | Output image, $ / M tokens |
|---|---|---|---|
| MiniMax-H3-Image | $2.30 | $1.50 | $3.80 |
| MiniMax-H3-Image-Highspeed | $2.30 | $3.00 | $7.60 |
The pricing page also gives example totals from MiniMax’s own tests on its China site, with default prompt enhancement. Those figures are “for reference only” and “do not represent fixed per-image prices.” For the standard model, MiniMax’s examples include about $0.02179 for a 1K text-to-image, $0.02333 for 2K, and $0.06879 for 4K. A 4K edit from one 1K reference is given as about $0.07312. The Highspeed examples are higher, including about $0.13001 for a 4K text-to-image. The older image-01 and image-01-live models are billed at $0.0035 per image. In its launch post, MiniMax also says there is “50% off 16MP+ generations during launch week.”
Licence
The weights are released under the MiniMax-H3-Image Community License Agreement, dated 10 November. The agreement applies worldwide except the European Union, the United Kingdom, the Republic of Korea and the United States, unless you obtain a separate licence from MiniMax. You also need prior written authorisation, requested through platform.minimax.io/license, if the annualised revenue of you and your affiliates exceeds US$1 million, or if you provide a hosted service to third parties. Commercial products that use the model must prominently display “MiniMax-H3-Image” on the user interface. MiniMax says it “claims no rights over the Outputs you generate.”
The prompt-expander weights are part of the same licence. MiniMax notes that MiniMax-H3-Image-PE is derived from Qwen3.8-27B, which is under the Apache 2.0 licence.
What testers have shown
MiniMax’s own model card includes official examples of photoreal photography, character work, Chinese and English text on a board, a game-map edit turned into a first-person view, and a living-room composition built from ten reference images. Those images are MiniMax’s demonstrations, not an independent benchmark. The card does not publish numeric leaderboard scores.
People with earlier Hailuo access had already been posting stills and edits. Mirochill wrote that the model seemed strong at following prompts and at using up to 16 reference images, but that with a simple prompt it lagged the competition, and that for everyday use he would still prefer GPT Image 2.6.
Prompt engineer underwood compared the same game-scene prompt on MiniMax-H3-Image and OpenAI’s GPT Image 2.5, and said MiniMax-H3-Image “looks a bit flatter to me.”
A separate community model, P-H3-Image from Pruna AI, already used MiniMax-H3 video weights for stills and edits with up to eight reference images. That’s a third-party derivative, not this official H3-Image release.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
