MiniMax releases H3-Image weights for 4K images in 10 seconds
The community license excludes the US, EU, UK and South Korea, and hosted services or revenue above $1 million need MiniMax's written approval.
Key takeaways
- MiniMax released the weights for H3-Image, an open-weight image model built on its H3 video model, on November 10, 2026.
- MiniMax says H3-Image generates a 4096x4096 (16 MP) image in under 10 seconds and takes up to 16 reference images in one request.
- The MiniMax-H3-Image Community License excludes the United States, the European Union, the United Kingdom and South Korea, and requires separate written authorization above $1 million in annual revenue or for any hosted service.
- MiniMax puts pricing at 50 to 70 percent of comparable mainstream offerings, with 50 percent off 4K and above requests during launch week.
MiniMax released H3-Image on November 10, 2026, an open-weight image model that generates a 4096x4096 image in under 10 seconds.
The company announced the release in a blog post and published the weights on GitHub and Hugging Face.
H3-Image handles text-to-image generation, reference-based generation and image editing in one checkpoint. It accepts up to 16 reference images in a single request. MiniMax says the model produces about 16 megapixels at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), directly and without an upscaler.
The model belongs to the H3 family that MiniMax built around its H3 video model, and it is an image model, not a video model. It reuses the H3 text encoder, a full Qwen3-VL encoder, and follows the H3-Base design with the audio branch removed. The diffusion transformer holds 33.1 billion parameters across 50 layers, and the published checkpoint is 68.65 GB in BF16. A full local run also needs the Qwen3-VL text encoder, which the image repository loads from the separate H3 video model rather than shipping itself.
The checkpoint is distilled to 6 sampling steps with no classifier-free guidance, so there is no negative prompt to write and the step count is fixed. ComfyUI added native support in a post dated November 10, with text-to-image and edit templates and partner nodes.
MiniMax says H3-Image runs at up to 5 times the speed of mainstream alternatives, and puts its Standard mode at 1.5 to 2 times as fast and its Highspeed mode at 3 to 4 times as fast. The company says pricing is 50 to 70 percent of comparable mainstream offerings, with 50 percent off 4K and above requests during launch week. Those figures are MiniMax's own claims.
The catch is the license. The MiniMax-H3-Image Community License, dated November 10, 2026, limits the grant of rights to an "Applicable Territory" that excludes the European Union, the United Kingdom, the Republic of Korea and the United States. A user whose annualized revenue passes $1 million, or who runs a hosted service for third parties, needs separate prior written authorization from MiniMax. Commercial products must also display "MiniMax-H3-Image" on their interface.
For creators in the four excluded regions, the weights are downloadable but the license grants no rights to use them without MiniMax's approval. Everyone else gets a 4K image model that can run locally on their own hardware.
H3-Image is also served through the MiniMax API Platform, MiniMax Code and Hailuo AI. To run it locally, download the checkpoint and the H3 text encoder from Hugging Face and load the ComfyUI template.
Sources
- minimax.io - release, 4K-in-10-seconds claim, 16 reference images, speed and pricing claims, availability
- github.com - 16 MP at every aspect ratio, 6 sampling steps, architecture, 68.65 GB checkpoint
- huggingface.co - weights, model zoo sizes, separate Qwen3-VL text encoder
- raw.githubusercontent.com - license date, excluded territories, $1M and hosted-service terms, UI display rule
- technode.com - independent report of the release and 16 reference images
- blog.comfy.org - ComfyUI support, 6-step sampling, no negative prompt