MiniMax releases H3-Image open weights with claimed 16-megapixel output

• H3-Image ships open weights and native 4096 × 4096 output, about 16.8 megapixels, in roughly 10 seconds according to MiniMax's launch announcement.
• Editing supports up to 16 reference images; WesRoth's overview notes that the API currently lacks masked edits and transparent backgrounds.
• Hosted access already spans fal, Runware and Vercel AI Gateway, as announced in MiniMax's fal post, its Runware announcement and Vercel's CLI example.

The model card fixes sampling at six steps and lists a separate 66.7 GB text-encoder download alongside the image checkpoint. The community license is the biggest catch: its standard grant excludes the US, EU, UK and South Korea.

MiniMax uses “4K” for an approximately 16-megapixel canvas across aspect ratios. Its model card describes direct generation without an upscaler and lists these native sizes:

• Square: 4096 × 4096.
• 16:9: 5472 × 3072, with the reverse dimensions for portrait.
• 21:9: 6272 × 2688.

The hosted API also accepts custom dimensions up to 20 million pixels, with aspect ratios between 1:3 and 3:1. Its area: "4K" setting specifies target pixel area; the server calculates dimensions from the requested aspect ratio.

Under the hood, the released model uses a 33.1-billion-parameter, 50-layer dense diffusion transformer derived from H3 with the audio branch removed. Text and reference images enter through H3's unchanged Qwen3-VL encoder; reference images also pass through the image VAE.

The checkpoint is distilled to six denoising evaluations without classifier-free guidance. Its local serving interface fixes the usual sampling controls:

• num_inference_steps: omitted or 6; other values return HTTP 400.
• guidance_scale: omitted or 1; other values return HTTP 400.
• negative_prompt: unsupported and rejected.

The VAE downsamples spatially by 16, and the transformer applies 2 × 2 latent patches, giving each transformer token a 32 × 32-pixel footprint.

The download inventory separates the image checkpoint from its text encoder:

• Image model repository: 68.65 GB, including the diffusion transformer and VAE.
• H3 text encoder: another 66.7 GB, loaded from the separate MiniMax-H3 repository and reusable by existing H3 installations.
• Optional prompt expander: 17.9 GB in NVFP4 format.

Those are download sizes, not GPU-memory requirements. MiniMax publishes deployment paths for SGLang, vLLM-Omni, diffusers and ComfyUI, including multi-GPU sequence-parallel configurations for SGLang and vLLM-Omni.

Native support also shipped in ComfyUI, alongside hosted Partner Nodes. In its editing graph, image_1 is the picture being edited and subsequent images are references; connecting the VAE anchors those references as latents as well as feeding them to the text encoder.

The editing endpoint accepts one to 16 reference images. MiniMax's generation guide specifies several additional constraints:

• Reference dimensions: 128 to 8192 pixels per dimension, with a maximum long-to-short-side ratio of 4:1.
• Request body: at most 64 MiB. Base64 encoding expands the original file data by roughly one-third.
• Prompt rewriting: image edits always apply prompt enhancement; the text actually used is returned in data[].revised_prompt.
• Unsupported operations: masked edits, transparent backgrounds and preview images.

Text-to-image also enables prompt enhancement by default, but permits disabling it with disable_prompt_enhanced: true. The local checkpoint expects a detailed English description; MiniMax supplies a separate prompt-expansion model for local workflows.

In early access, ProperPrompter asked for two characters “exactly as they appear” in a two-reference prompt that also requested “mild compression artifacts.”

MiniMax's API exposes MiniMax-H3-Image and MiniMax-H3-Image-Highspeed with identical parameters and limits. Generation uses POST /v1/images/generations; editing uses POST /v1/images/edits.

MiniMax Code and Hailuo AI also offer access. The launch announcement offers 50% off standard pricing for 4K-and-above generation requests during launch week.

Distribution includes fal, while Vercel's announcement supplies an AI CLI invocation and MiniMax's Runware post emphasizes multi-reference workflows.

MiniMax also promoted a HeyGen MCP contest for HeyGen Video 1, built on the H3 video model, with $5,000 in MiniMax API credits and a December 1 deadline.

The downloadable weights use MiniMax's custom Community License Agreement, dated November 10, 2026. Its restrictions extend beyond downloading or modifying the checkpoint:

• Territory: the standard grant excludes the EU, UK, South Korea and US. The restrictions also cover use or display of outputs outside the permitted territory; MiniMax invites separate licensing requests from excluded regions.
• Commercial authorization: prior written authorization is required if annualized revenue of the licensee or its affiliates exceeds $1 million, or if it provides a hosted service to third parties. Hosting triggers the requirement independently of revenue.
• Model training: outputs cannot improve other AI models, except H3-Image or its model derivatives.
• Branding: commercial products or services using the model must prominently display “MiniMax-H3-Image” in their interface.

MiniMax claims no rights over generated outputs, while leaving responsibility for their subsequent use with the operator and users.

The text-to-image endpoint waits up to 600 seconds, according to its API reference. Generation continues after a client timeout or disconnect, and successful images are still charged.

Requests are non-idempotent: retrying generates another image and incurs another charge. MiniMax's OpenAI SDK example sets timeout=600 and max_retries=0.

Batch responses have their own failure semantics, documented in the generation guide:

• A request can generate up to 10 images.
• HTTP 200 means at least one image succeeded, not that the entire batch succeeded.
• Failed or safety-withheld entries contain error; usage.generated_images records the successful count.
• Streaming emits completed images individually through SSE, in completion order, identified by image_index.