MiniMax on Monday released H3-Image, an open-weight image generation model built on its H3 foundation, and published the weights on Hugging Face at the same time as the launch. The model handles text-to-image generation, reference-based generation and image editing in a single system, continuing the Chinese lab's pattern of shipping frontier-adjacent capability and opening it immediately.
The headline claims, all company-reported: H3-Image generates a full 4K image — 4096 x 4096, roughly 16 megapixels — in under 10 seconds, at up to five times the speed of mainstream alternatives. The model accepts up to 16 reference images, a practical ceiling for e-commerce and brand-design workflows where multi-angle product photos, brand palettes and model poses all need to stay consistent. The model card lists an effective resolution range of 4 to 20 megapixels, and a 21:9 frame maps to 6272 x 2688 without a separate upscaling model.
The architecture details are unusually complete for a launch-day release. The backbone is a 33.1-billion-parameter diffusion Transformer with 50 layers and 56 attention heads, kept in BF16. The text encoder is the Qwen3-VL encoder from MiniMax's H3 video model, used unchanged. The image VAE is new: a CNN encoder paired with a 24-layer 2D ViT decoder, with 16x spatial downsampling so each latent token covers 32 x 32 pixels.
The speed comes mainly from distillation. The released checkpoint is compressed to six sampling steps and runs without classifier-free guidance — against a conventional diffusion pipeline that often needs 20 to 30 steps, with CFG doubling the forward passes per step. Removing both is what turns minute-scale 4K generation into a ten-second operation, and it is the same recipe MiniMax applied when the community distilled H3's video checkpoints.
MiniMax also open-sourced a companion: H3-Image-PE, a 27-billion-parameter prompt-expansion model fine-tuned from Qwen3.8-27B that turns a short request into a detailed scene description and recommends an aspect ratio. Deployment support landed on day one across SGLang, vLLM-Omni, Diffusers and ComfyUI, with weights on both Hugging Face and ModelScope.
On the API side, the company says Standard mode runs 1.5 to 2 times faster than mainstream alternatives and Highspeed mode 3 to 4 times faster, with pricing set at 50 to 70 percent of comparable offerings. Requests at 16MP and above are half price for launch week. The model is available through the MiniMax platform, MiniMax Code and Hailuo AI. One caveat for commercial users: H3-Image ships under a community license, and some uses require written authorization through an application process.
The boldest claim — state-of-the-art prompt adherence among open-weight models — rests on MiniMax's own account. The model card includes no specific benchmark scores, and independent evaluations have not yet landed. Image editing is also the least finished part of the release: only a prompt template has been published, with the repository marking that portion as still in development.
The lineage explains the strategy. In August, MiniMax opened the weights of H3's video model — by several accounts the first open model to reach SOTA-level video generation — and the community response, including derivative models like Pika's H3 Flex, validated the open-first playbook. H3-Image applies the same sequence to stills: ship video first, split out images, release weights immediately, and let third-party benchmarks — eventually — do the marketing.
Comments (0)
Log in to join the discussion
Log InNo comments yet