What it does
MiniMax H3-Image is the image member of MiniMax's H3 family, released on Hugging Face in November 2026. It is the H3 video model's diffusion transformer with the audio layers removed, reusing the H3 text encoder, and one 33.1B-parameter checkpoint does both text-to-image and editing with up to 16 reference images. Images come out at about 16 megapixels (4096 × 4096 at 1:1, up to 6272 × 2688 at 21:9) with no upscaler, in 6 sampling steps. The weights are downloadable under MiniMax's own community license, and the same model is sold through the MiniMax API and used in the Hailuo AI app.
Official source 1 ↗Features
- 16-megapixel output. Generates about 16 MP directly at every aspect ratio, from 4096 × 4096 to 6272 × 2688, with a working range of about 4–20 MP and no upscaler. Official source 1 ↗
- Generation and editing in one model. The same checkpoint creates images from text and edits one or more reference images, up to 16 at once; MiniMax's examples compose an empty room and nine furniture pieces into one living room. Official source 1 ↗
- Six-step sampling. The checkpoint is distilled to 6 steps with no classifier-free guidance, which MiniMax says keeps 16 MP generation practical. Official source 1 ↗
- Built on MiniMax H3. It reuses the H3 video model's architecture without the audio layers and loads the H3 text encoder unchanged, so an existing H3 download is not fetched again. Official source 1 ↗
- Prompt expansion. The model expects a detailed English description. The API expands short prompts in any language with H3-Image-Context-IR, and the optional H3-Image-PE expander can run locally. Official source 1 ↗ · Official source 2 ↗
- Text and reasoning in images. MiniMax's showcase includes logic puzzles in Chinese and English solved on a board, and a game map turned into a first-person view with a minimap. Official source 1 ↗
Pricing & free limits
Licensed worldwide except the US, EU, UK and South Korea. Written authorization is needed above $1M annual revenue or to offer it as a hosted service. Official source 1 ↗ · Official source 2 ↗
Example costs: $0.02179 for a 1K text-to-image result, $0.06879 for 4K, $0.07914 for a 4K edit with three references. Official source 1 ↗
Faster variant; a 4K text-to-image example costs $0.13001. Official source 1 ↗
Billing details
API charges follow actual token usage; MiniMax's per-image figures are rounded examples, not fixed prices, and text tokens can make costs vary. Official source 1 ↗
Until November 17, 2026 at 16:00 UTC, requests whose final image is at least 16.6 million pixels cost half. Official source 1 ↗
Prompt expansion through H3-Image-Context-IR is billed separately at $2.30 per million input tokens. Official source 1 ↗
Free access & limits
Free access and its limits were not established by the checked official sources.
Best for & limitations
Producing print-size or high-detail images such as posters, concept art and product scenes without a separate upscaling step. Official source 1 ↗
Composing a scene from many reference images, for example placing a set of products or furniture into one room. Official source 1 ↗
Running a large image model on your own GPUs where the licence territory and revenue terms fit. Official source 1 ↗ · Official source 2 ↗
- Not licensed in the US, EU, UK or Korea. The community license excludes those regions entirely, including use of the outputs there; people in them need a separate licence from MiniMax. Official source 1 ↗
- Commercial strings attached. Above $1M annual revenue or for any hosted service you need MiniMax's prior written authorization, and commercial products must display "MiniMax-H3-Image" in their interface. Official source 1 ↗
- Outputs can't train other models. The license forbids using the model or its outputs to improve any other AI model (other than H3-Image and its own derivatives). Official source 1 ↗
- Very large to self-host. The DiT alone is 68.65 GB in BF16 and the H3 text encoder adds 66.7 GB; MiniMax's deployment examples target data-center GPUs. Official source 1 ↗
- Fixed sampling, no negative prompts. Steps and guidance are fixed (other values return HTTP 400) and negative prompts are rejected; short or non-English prompts need the expander first. Official source 1 ↗
- Safeguards required for public services. Anyone offering generation to others must put reasonable abuse safeguards in place before launch and keep testing and reviewing them. Official source 1 ↗
Submit a tool
