{"id":"minimax-h3-image","slug":"minimax-h3-image","name":"MiniMax H3-Image","orgId":"minimax","family":"minimax-h3-image","releaseDate":null,"availability":"ga","openWeights":true,"license":"MiniMax-H3-Image Community License Agreement (license date November 10, 2026)","paramsTotal":null,"paramsSource":null,"contextWindow":null,"knowledgeCutoff":null,"modelCardUrl":"https://x.com/MiniMax_AI/status/2120179737695826420","hfRepo":null,"modalities":["image"],"aliases":["MiniMax H3-Image","minimax-h3-image","minimax h3 image","H3-Image","h3 image","MiniMaxAI/MiniMax-H3-Image","MiniMax-H3-Image"],"aliasNote":null,"origin":"tracked","tracked":{"trackedSince":"2026-11-10","announced":"2026-11-10","blurb":"MiniMax's open-weight image generation and editing model in the H3 family: a 33.1B-parameter diffusion transformer that writes about 16-megapixel images natively (4096×4096 up to 6272×2688) and edits with up to 16 reference images. Distilled to 6 sampling steps with no classifier-free guidance. Available through the MiniMax API (model MiniMax-H3-Image), MiniMax Code and Hailuo AI, with weights on Hugging Face under the MiniMax-H3-Image Community License.","api":"live","apiModelId":"MiniMax-H3-Image","openWeights":"yes","params":{"text":"33.1B (dense diffusion transformer); text encoder is the unchanged MiniMax-H3 / Qwen3-VL encoder, loaded separately","source":"MiniMax-H3-Image Hugging Face model card"},"facts":[{"label":"Type","value":"Unified text-to-image and image-editing checkpoint (H3-Base design with the audio branch removed). Text and references go through the H3 text encoder and image VAE into a single-stream transformer"},{"label":"Output size","value":"About 16 megapixels natively, without an upscaler: 4096×4096 (1:1) up to 6272×2688 (21:9). The card's working range is about 4 MP to 20 MP"},{"label":"References","value":"Up to 16 reference images for editing (MiniMax announcement and model card)"},{"label":"Sampling","value":"Distilled to 6 sampling steps (6 NFEs); classifier-free guidance is not used (guidance_scale must be 1 on the served API)"},{"label":"Architecture","value":"Dense 50-layer DiT (hidden size 5376, 56 heads, FFN 14336) with 3D multimodal RoPE, same as the H3-Omni transformer; DiT weights BF16, with embedding, timestep and output layers in FP32. Image VAE: 16× spatial downsampling, 24 latent channels"},{"label":"Prompt expander","value":"Optional MiniMax-H3-Image-PE (NVFP4, 17.9 GB) for local prompt expansion; the MiniMax API uses H3-Image-Context-IR. Not a separate page"},{"label":"Local runtimes","value":"SGLang, vLLM-Omni, Diffusers and ComfyUI (model card). The text encoder is not in this repo; load MiniMaxAI/MiniMax-H3 text_encoder/"},{"label":"Hosted access","value":"MiniMax API (model MiniMax-H3-Image; text-to-image, image editing, Context-IR), MiniMax Code and Hailuo AI (MiniMax announcement). Launch-week 50% off 16MP+ generations (MiniMax on X)"},{"label":"Weights","value":"Public, ungated Hugging Face repo MiniMaxAI/MiniMax-H3-Image (BF16 DiT about 66.28 GB plus VAE; text encoder downloaded from MiniMax-H3)"},{"label":"License","value":"MiniMax-H3-Image Community License Agreement (license date November 10, 2026). Applicable territory is worldwide excluding the EU, UK, South Korea and the United States unless MiniMax grants a separate license. Prior written authorization is required if annualized revenue of the licensee and affiliates exceeds USD 1 million, or if hosting the model as a service for third parties. Commercial products must display “MiniMax-H3-Image” on the user interface"},{"label":"Code","value":"https://github.com/MiniMax-AI/MiniMax-H3-Image (first public commit November 10, 2026)"},{"label":"Related","value":"Separate from MiniMax-H3 (the video model), P-H3-Image, Hailuo video models, MiniMax image-01, and MiniMax text models such as M3 / M3.1 Flash"}],"claims":["Lab claim: generates 4K / 16-megapixel images in 10 seconds (MiniMax announcement on X).","Lab claim: the diffusion transformer has 33.1B parameters; images of about 16 megapixels are produced directly, without an upscaler (MiniMax Hugging Face card).","Lab claim: delivers the strongest instruction following among open-source models and supports up to 16 reference images (MiniMax Hugging Face card)."],"pitch":"an open-weight image model for 16-megapixel generation and editing, with up to 16 reference images, in about 10 seconds","sources":[{"kind":"lab","title":"MiniMax on X: MiniMax H3-Image","url":"https://x.com/MiniMax_AI/status/2120179737695826420","date":"2026-11-10"},{"kind":"lab","title":"MiniMax platform release notes (Nov. 10, 2026)","url":"https://platform.minimax.io/docs/release-notes/models","date":"2026-11-10"},{"kind":"lab","title":"MiniMax image generation docs","url":"https://platform.minimax.io/docs/guides/image-generation","date":"2026-11-10"},{"kind":"repo","title":"MiniMaxAI/MiniMax-H3-Image on Hugging Face","url":"https://huggingface.co/MiniMaxAI/MiniMax-H3-Image","date":"2026-11-10"},{"kind":"repo","title":"MiniMax-AI/MiniMax-H3-Image on GitHub","url":"https://github.com/MiniMax-AI/MiniMax-H3-Image","date":"2026-11-10"}],"firstBenchmark":{"date":"2026-11-13","benchmarkId":"lmarena-t2i"}},"listedSince":"2026-11-12","organization":{"id":"minimax","name":"MiniMax","slug":"minimax","country":null,"url":null,"color":"#A39E8F","mark":"MI"},"price":null,"answer":{"word":"Yes.","scope":"Yes on Image generation (open weights); Image edit (open weights). Not the overall leader.","lead":"As of Nov 22, 2026, MiniMax H3-Image is tied for state of the art in Image generation (since Nov 17, 2026; sources LMArena Text-to-Image; methodology v1.0). MiniMax H3-Image is tied for SOTA in Image generation with Ideogram 4.0, within the margin of error, since 2026-11-17.","confidence":"medium"},"status":"BENCHMARKED","verdicts":[{"id":"v_overall_image-gen_minimax-h3-image_2026-11-22_open_RANKED","model":"minimax-h3-image","model_name":"MiniMax H3-Image","category":"image-gen","category_name":"Image generation","slice":"overall","state":"RANKED","rank":23,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":27.84810126582279,"confidence":"medium","valid_from":"2026-11-22","valid_to":null,"configuration":null,"rationale":"Ranked #23 in Image generation.","safeguard_flags":[],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"#23 in Image generation"},{"id":"v_open-weights_image-gen_minimax-h3-image_2026-11-17_open_CO_SOTA","model":"minimax-h3-image","model_name":"MiniMax H3-Image","category":"image-gen","category_name":"Image generation","slice":"open-weights","state":"CO_SOTA","rank":1,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":0,"confidence":"medium","valid_from":"2026-11-17","valid_to":null,"configuration":null,"rationale":"MiniMax H3-Image is tied for SOTA in Image generation with Ideogram 4.0, within the margin of error, since 2026-11-17.","safeguard_flags":["two_signal_confirmed"],"co_sota_with":["ideogram-4-0"],"challengers":[],"methodology":"1.0","label":"Co-SOTA · Image generation (open weights)"},{"id":"v_overall_image-edit_minimax-h3-image_2026-11-22_open_RANKED","model":"minimax-h3-image","model_name":"MiniMax H3-Image","category":"image-edit","category_name":"Image edit","slice":"overall","state":"RANKED","rank":23,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":38.59649122807018,"confidence":"medium","valid_from":"2026-11-22","valid_to":null,"configuration":null,"rationale":"Ranked #23 in Image edit.","safeguard_flags":[],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"#23 in Image edit"},{"id":"v_open-weights_image-edit_minimax-h3-image_2026-11-17_open_SOTA","model":"minimax-h3-image","model_name":"MiniMax H3-Image","category":"image-edit","category_name":"Image edit","slice":"open-weights","state":"SOTA","rank":1,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":0,"confidence":"medium","valid_from":"2026-11-17","valid_to":null,"configuration":null,"rationale":"MiniMax H3-Image is SOTA in Image edit since 2026-11-17. No eligible model is credibly ahead on the weighted primary evidence.","safeguard_flags":["two_signal_confirmed"],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"SOTA · Image edit (open weights)"}]}