MiniMax Group Inc (hereafter MiniMax) says MiniMax H3-Image (hereafter H3-Image) produces a roughly 16-megapixel image — 4,096 × 4,096 at 1:1 — in under 10 seconds, which the company describes as about five times faster than the market norm for that resolution3. MiniMax set API pricing at 50–70% of what it calls comparable tools, and says standard-mode speed is 1.5–2× that of major tools, rising to 3–4× in a Highspeed tier4. No independent verification of those performance or pricing comparisons was provided5. The model went live on November 10 with open weights on Hugging Face Inc. (hereafter Hugging Face) for text-to-image generation and image editing1,7.
H3-Image carries a community license that excludes use in the US, the EU, the UK, and South Korea, a restriction that limits access for a large share of the model's potential developer base. Larger companies need written authorization from MiniMax to use the weights at all.
Pricing and promotional window
MiniMax's API charges per token rather than per image. Standard-tier rates are $2.30 per million input text tokens, $1.50 per million input image tokens, and $3.80 per million output image tokens; a Highspeed tier keeps the input rates and charges $7.60 for output8,9. At those rates, a 1,024 × 1,024 text-to-image request costs approximately $0.022 and a 4,096 × 4,096 request costs approximately $0.069 at list price. A 50% discount on requests delivering at least 16.6 million pixels runs until November 17 at 16:00 UTC.
Architecture and serving
H3-Image is a 33.1-billion-parameter diffusion transformer derived from MiniMax's H3 family, using the H3 design without the audio layers. H3 is MiniMax's video-generation model. The model is distilled to six sampling steps without classifier-free guidance; API requests with other step counts or guidance values return HTTP 400. A single checkpoint handles both text-to-image generation and editing with one or more reference images, accepting up to 16 references2.
Unlike typical image-generation systems that create a low-resolution image and then upscale it, H3-Image directly outputs high resolution. Output sizes reach up to 6,272 × 2,688 without an upscaler.
The BF16 checkpoint is 68.65 GB, with an optional prompt-expansion model called MiniMax-H3-Image-PE at 17.9 GB. Full local functionality depends on a separate text encoder, and a single 80 GB card will hold the transformer but not the transformer plus the encoder at full precision. The model card recommends SGLang, vLLM-Omni, Diffusers, and ComfyUI for local serving, and lists Nvidia Corp. (hereafter Nvidia) and AMD accelerators. ComfyUI supported H3-Image at release.
MiniMax released examples including photo-realistic images, 3D game concept images, and images that include text. The company calls the model's instruction following the strongest among open-source image models and says it avoids the familiar "AI look". Those are the company's own claims; no independent comparison was provided.