MiniMax releases H3-Image: 33.1B-parameter model generates 16-megapixel images in 6 sampling steps
MiniMax has released MiniMax-H3-Image, a 33.1B-parameter diffusion transformer that handles text-to-image and image editing in one checkpoint. It generates images of about 16 megapixels directly, without an upscaler, using 6 sampling steps and no classifier-free guidance.
MiniMax has released MiniMax-H3-Image, a 33.1B-parameter diffusion transformer from its H3 family that generates images of about 16 megapixels directly, without an upscaler. The model is distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance. Weights are published on Hugging Face under MiniMaxAI/MiniMax-H3-Image. API pricing is not yet disclosed.
Key specifications
- Task coverage: text-to-image and editing of one or more reference images in a single checkpoint. It supports up to 16 reference images.
- Output size: about 16 MP at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9). Other listed 16 MP sizes include 4736x3552 (4:3), 5024x3360 (3:2) and 5472x3072 (16:9), plus their portrait counterparts.
- Working range: about 4 MP to 20 MP, with the best results generally near 16 MP.
- Sampling: 6 steps, no CFG. Other values for
num_inference_stepsorguidance_scalereturn HTTP 400, andnegative_promptis rejected. - Prompt length: prompts are used as given and truncated to 4096 tokens.
Architecture
The diffusion transformer is a dense single-stream model with 50 layers, hidden size 5376, 56 attention heads, FFN size 14336 and 3D multimodal RoPE. It is the H3 architecture with the audio branch removed. Weights are BF16, with the embedding, timestep and output layers in FP32.
The text encoder is the full Qwen3-VL text encoder from MiniMax-H3, unchanged, and the DiT reads hidden states from its 50th layer. Text prompts and reference images are encoded by that encoder, and reference images are also encoded by the image VAE. All tokens form one packed sequence.
The image VAE uses a CNN encoder and a 2D ViT decoder (24 layers) with 16x spatial downsampling and 24 latent channels. With a 2x2 latent patch, each transformer token covers 32x32 pixels.
Files and deployment
The MiniMax-H3-Image model zoo entry lists the DiT at 68.65 GB in BF16. The repository holds the transformer (66.28 GB), the VAE (2.34 GB), the tokenizer and processor (23 MB), and per-step output heads with the sampling schedule (12 MB). The text encoder (66.7 GB) is not in this repository. It loads from the text_encoder/ folder of MiniMaxAI/MiniMax-H3 and is not downloaded again if already cached.
An optional prompt expander, MiniMax-H3-Image-PE, is listed at 17.9 GB in NVFP4. The model expects a detailed English description of the whole image. According to MiniMax, its API converts a short prompt in any language into such a description using H3-Image-Context-IR before generation.
Recommended inference frameworks are SGLang, vLLM-Omni, diffusers and ComfyUI. The vLLM-Omni recipe covers NVIDIA H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X, among others. Both frameworks can split a request across several GPUs with Ulysses sequence parallelism, for example on 4 GPUs. The server exposes /v1/images/generations for text-to-image and /v1/images/edits for editing, and both return a PNG in data[0].b64_json.
Hosted access is available through the MiniMax API (platform.minimax.io globally, platform.minimax.cn in China), the Hailuo web app (hailuoai.video and hailuoai.com) and the desktop agent (agent.minimax.io and agent.minimax.cn).
Claims
MiniMax describes the model as state-of-the-art and says it delivers the strongest instruction following among open-source models. It also claims realistic output that goes beyond the "AI look" and a computational cost that keeps 16 MP generation practical. The model page lists no benchmark scores.
What this means
MiniMax-H3-Image pairs a high native resolution with a short, fixed sampling schedule. The 6-step, no-CFG distillation is the stated basis for its claim that 16 MP generation is practical, and the API enforces it by rejecting other step and guidance values. Combining generation and editing in one checkpoint, with up to 16 reference images, means one set of weights covers both workflows. Reusing the MiniMax-H3 text encoder unchanged means anyone already holding MiniMax-H3 avoids a second download of that component. Teams planning local deployment should budget for the DiT and the separate text encoder, since their listed sizes are 68.65 GB and 66.7 GB. MiniMax's instruction-following claim is untested by any benchmark on the model page.
Related Articles
Underdog Woof, a 40.6 MB two-voice TTS model, claims 1.42% word error rate on a CPU
Underdog Woof is an English text-to-speech model with two voices, Juno and Mira, packaged as a single 40.6 MB file that runs on a CPU with numpy. Underdog reports a 1.42% word error rate for Mira on Gradium TTS Hard Cases, against 1.89% for Fish Audio Drama 3, a 41.5 GB download.
ByteDance's Seedream 5.5 Holds Up Through 100 Edit Rounds in QbitAI Test; 2K Edit Costs About 0.28 Yuan
ByteDance's Seedream 5.5 stayed more faithful than GPT-Image 2.6 in QbitAI's 100-round sequential editing test, though some distortion appeared in later rounds. According to official information, a 2K square image edit via API costs about 0.28 yuan, under half the price of Seedream 5.0 Pro.
Bytedance Seedream 5.5 claims pixel-stable image editing across up to 100 rounds
Bytedance Seed has introduced Seedream 5.5, an image generation model for professional production workflows. According to Bytedance, it keeps unchanged areas pixel-identical across up to 100 editing rounds and generates native RGBA images with transparent backgrounds.
Video Rebirth releases HyperFlow H3 Flash, a 14B few-step video and audio model compressed from MiniMax-H3
Video Rebirth has published HyperFlow H3 Flash, a compressed, few-step version of MiniMax-H3 with 13.7B to 13.9B parameters. It generates 5.2 to 15 second videos at 24 fps with synchronized 32 kHz stereo audio, using 4, 6 or 8 sampling steps without classifier-free guidance.
Comments
Loading...