MiniMax H3-Image is an open-weight image model for text-to-image and multi-reference editing that MiniMax says can make native 4K images in under 10 seconds.

What it is: one checkpoint for generate and edit

MiniMax announced H3-Image on November 11, 2026, with weights dated November 10. It is built on the MiniMax H3 family with the audio layers removed. Generation and editing share one checkpoint: text alone makes a new image, and a prompt plus reference images edits or composes a scene. The model card says it can take up to 16 reference images.

Native high resolution: about 16 megapixels without an upscaler. Common sizes include 4096×4096 (1:1), 5472×3072 (16:9) and 6272×2688 (21:9). The working range is about 4 MP to 20 MP, with best results around 16 MP.
Six sampling steps: the checkpoint is distilled to 6 steps with no classifier-free guidance, so there is no negative prompt to tune. MiniMax's blog says a full 4K image can finish in under 10 seconds, and up to 5× the speed of mainstream alternatives. That is MiniMax's claim, not a Seailife timing.
Architecture numbers from the model card: a 33.1B-parameter diffusion transformer, BF16 DiT weights of about 68.65 GB, and the MiniMax-H3 Qwen3-VL text encoder unchanged. An optional prompt expander, MiniMax-H3-Image-PE, is a separate 17.9 GB download.

Where to run it

Weights:Hugging Face, with a mirror on ModelScope. The DiT and VAE live in the H3-Image repo; the text encoder is pulled from MiniMax-H3. If you already have H3, that encoder is not downloaded again.
Local runtimes named in the README: SGLang, vLLM-Omni, Hugging Face Diffusers, LightX2V and ComfyUI (native support noted by Comfy on November 10).
Hosted: the MiniMax API, MiniMax Code, and Hailuo AI on the web. There is also a Highspeed API model, MiniMax-H3-Image-Highspeed, with the same parameters as the standard model.
GitHub:MiniMax-AI/MiniMax-H3-Image points to the same release. On November 12 it had about 265 stars and 13 forks.

API shape and hard limits

Text-to-image is POST /v1/images/generations. Image edit is POST /v1/images/edits with 1 to 16 reference images. Both can stream results image by image. A request can return 1 to 10 images.
Output size is auto with an aspect ratio and area level (1K, 2K, or default 4K), or an exact WIDTHxHEIGHT. The area must sit between 655,360 and 20,000,000 pixels, with aspect ratio between 1:3 and 3:1.
Prompt length is up to 32,000 characters. With prompt enhancement off, only about the first 4,096 tokens take effect. MiniMax expects a detailed English description for local runs; the API can expand a short prompt in any language with H3-Image-Context-IR, and the PE model can do that locally.
The API docs say it does not support preview images, masked edits or transparent backgrounds.

API pricing, as listed on November 12

Standard MiniMax-H3-Image: $2.30 per million text-input tokens, $1.50 per million image-input tokens, $3.80 per million image-output tokens.
Highspeed: the same text rate, then $3.00 and $7.60 for image in and out.
MiniMax's example for a standard text-to-image 4K (4096×4096) run is about $0.06879, or $0.03439 with the launch discount. The 50% off for delivered images at or above 16,600,000 pixels runs until November 17, 2026, 16:00 UTC. Example totals change with prompt length and reference images.
The research blog also says Standard is 1.5–2× as fast as mainstream alternatives, Highspeed 3–4×, and pricing 50–70% of comparable offerings. Those speed and price comparisons are MiniMax's, not checked here.

License and what this page is not

Weights use the MiniMax-H3-Image Community License. The Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States. People in those places who want to deploy the model are told to contact MiniMax about a separate license. Hosted API and Hailuo access are separate from that self-host rule.
This is not the H3 video model already listed as MiniMax H3, and not Pika H3 Flex. It is the image checkpoint only.
Seailife did not download the weights, call the API or compare image quality. Speed, realism and prompt-adherence claims above come from MiniMax's blog, model card and API docs. The logo is Hailuo's mark from the product page; the product image is the official showcase grid from Hugging Face.

What it is: one checkpoint for generate and edit

Native high resolution: about 16 megapixels without an upscaler. Common sizes include 4096×4096 (1:1), 5472×3072 (16:9) and 6272×2688 (21:9). The working range is about 4 MP to 20 MP, with best results around 16 MP.

Six sampling steps: the checkpoint is distilled to 6 steps with no classifier-free guidance, so there is no negative prompt to tune. MiniMax's blog says a full 4K image can finish in under 10 seconds, and up to 5× the speed of mainstream alternatives. That is MiniMax's claim, not a Seailife timing.

Architecture numbers from the model card: a 33.1B-parameter diffusion transformer, BF16 DiT weights of about 68.65 GB, and the MiniMax-H3 Qwen3-VL text encoder unchanged. An optional prompt expander, MiniMax-H3-Image-PE, is a separate 17.9 GB download.

Where to run it

Weights:Hugging Face, with a mirror on ModelScope. The DiT and VAE live in the H3-Image repo; the text encoder is pulled from MiniMax-H3. If you already have H3, that encoder is not downloaded again.

Local runtimes named in the README: SGLang, vLLM-Omni, Hugging Face Diffusers, LightX2V and ComfyUI (native support noted by Comfy on November 10).

Hosted: the MiniMax API, MiniMax Code, and Hailuo AI on the web. There is also a Highspeed API model, MiniMax-H3-Image-Highspeed, with the same parameters as the standard model.

GitHub:MiniMax-AI/MiniMax-H3-Image points to the same release. On November 12 it had about 265 stars and 13 forks.

API shape and hard limits

Text-to-image is POST /v1/images/generations. Image edit is POST /v1/images/edits with 1 to 16 reference images. Both can stream results image by image. A request can return 1 to 10 images.

Output size is auto with an aspect ratio and area level (1K, 2K, or default 4K), or an exact WIDTHxHEIGHT. The area must sit between 655,360 and 20,000,000 pixels, with aspect ratio between 1:3 and 3:1.

Prompt length is up to 32,000 characters. With prompt enhancement off, only about the first 4,096 tokens take effect. MiniMax expects a detailed English description for local runs; the API can expand a short prompt in any language with H3-Image-Context-IR, and the PE model can do that locally.

The API docs say it does not support preview images, masked edits or transparent backgrounds.

API pricing, as listed on November 12

Standard MiniMax-H3-Image: $2.30 per million text-input tokens, $1.50 per million image-input tokens, $3.80 per million image-output tokens.

Highspeed: the same text rate, then $3.00 and $7.60 for image in and out.

MiniMax's example for a standard text-to-image 4K (4096×4096) run is about $0.06879, or $0.03439 with the launch discount. The 50% off for delivered images at or above 16,600,000 pixels runs until November 17, 2026, 16:00 UTC. Example totals change with prompt length and reference images.

The research blog also says Standard is 1.5–2× as fast as mainstream alternatives, Highspeed 3–4×, and pricing 50–70% of comparable offerings. Those speed and price comparisons are MiniMax's, not checked here.

License and what this page is not

Weights use the MiniMax-H3-Image Community License. The Applicable Territory is worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States. People in those places who want to deploy the model are told to contact MiniMax about a separate license. Hosted API and Hailuo access are separate from that self-host rule.

This is not the H3 video model already listed as MiniMax H3, and not Pika H3 Flex. It is the image checkpoint only.

Seailife did not download the weights, call the API or compare image quality. Speed, realism and prompt-adherence claims above come from MiniMax's blog, model card and API docs. The logo is Hailuo's mark from the product page; the product image is the official showcase grid from Hugging Face.

MiniMax H3-Image

What it is: one checkpoint for generate and edit

Where to run it

API shape and hard limits

API pricing, as listed on November 12

License and what this page is not

Comments

MiniMax H3-Image

What it is: one checkpoint for generate and edit

Where to run it

API shape and hard limits

API pricing, as listed on November 12

License and what this page is not

Comments