MiniMax releases 16-megapixel open-weight image model with unified editing

MiniMax has released MiniMax-H3-Image, a new open-weight model for high-resolution image generation and editing. Published on November 10, 2026, this release marks a significant step in open-source visual AI by combining text-to-image creation and reference-based editing into a single architecture. The model targets developers and creators who need precise control over large-scale visual outputs without relying on proprietary APIs.

What happened

The MiniMax team introduced MiniMax-H3-Image as part of their H3 family of models. Unlike many previous open-source generators that focus solely on creating images from text, this model unifies generation and editing in one checkpoint. It accepts text prompts for new creations and can also process up to 16 reference images to guide edits or compose complex scenes. This dual capability allows users to maintain consistent characters or styles across different visual tasks while keeping the workflow within a single system.

A defining feature of this release is its native support for high-resolution outputs. The model generates images at approximately 16 megapixels directly, such as 4096x4096 pixels for square formats or 6272x2688 for wide cinematic ratios. It achieves this without requiring a separate upscaling stage, which often introduces artifacts or blurs details in other pipelines. By handling these dimensions natively, the model preserves intricate textures and realistic lighting that are typically lost in lower-resolution generations followed by enlargement.

The release includes detailed specifications for deployment and inference. The main checkpoint weighs 68.65 GB in BF16 precision, while an optional prompt expander model adds another 17.9 GB. The team provided integration guides for popular inference frameworks like SGLang, vLLM-Omni, and diffusers. They also shared example scripts that demonstrate how to replicate the quality of their hosted API using local hardware, ensuring that developers can test and deploy the model independently.

How it works

MiniMax-H3-Image is built on a diffusion transformer (DiT) architecture derived from the H3-Base design, with the audio processing layers removed. The core engine contains 33.1 billion parameters arranged in 50 dense transformer layers. It uses a hidden size of 5376 and 56 attention heads, employing 3D multimodal rotary positional embeddings to handle spatial and temporal data efficiently. This structure allows the model to process text and visual information in a single stream, predicting the latent representation of the target image directly.

The input pipeline combines text and visual cues through specific encoders. Text prompts are processed by the unchanged Qwen3-VL text encoder from the broader MiniMax-H3 suite, with the DiT reading hidden states from the 50th layer. Reference images are encoded using a custom image VAE that features a CNN encoder and a 2D ViT decoder with 24 layers. This VAE downsamples spatial dimensions by a factor of 16, resulting in 24 latent channels. The DiT then patchifies these latents with a 2x2 patch size, meaning each transformer token corresponds to a 32x32 pixel area in the final output.

To achieve fast inference, the model is distilled to require only six sampling steps, known as noise function evaluations (NFEs). It operates without classifier-free guidance, which simplifies the sampling process and reduces computational overhead. The weights are stored in BF16 precision, while the embedding, timestep, and output layers use FP32 for stability. This configuration enables the model to run on various accelerators, including NVIDIA H200 and AMD Instinct MI350X GPUs, with options for sequence parallelism to distribute loads across multiple devices.

Key details

• Resolution: Generates native 16-megapixel images (e.g., 4096x4096) without upscalers.
• Speed: Distilled to just 6 sampling steps with no classifier-free guidance required.
• Architecture: 33.1 billion parameter diffusion transformer based on H3-Base.
• Inputs: Supports text-to-image and image editing with up to 16 reference images.
• Size: Main checkpoint is 68.65 GB; optional prompt expander is 17.9 GB.
• License: Released under the MiniMax-H3-Image Community License Agreement.

Why it matters

For software engineers building creative tools, the shift to native 16-megapixel generation removes a major bottleneck in production pipelines. Traditional workflows often generate small images and rely on separate upscaling models, which can degrade text legibility and fine details. By producing high-resolution outputs directly, MiniMax-H3-Image ensures that elements like small text in logic puzzles or intricate textures in environment concepts remain sharp. This reliability is crucial for professional applications where visual fidelity cannot be compromised by post-processing artifacts.

The unified approach to generation and editing also simplifies application logic. Developers no longer need to maintain separate models for creating new assets and modifying existing ones. The ability to ingest up to 16 reference images enables complex composition tasks, such as arranging furniture in a room or maintaining character consistency across different scenes. This flexibility supports more sophisticated user interactions, allowing creators to iterate on designs with greater precision and less manual intervention.

Furthermore, the open-weight nature of the release empowers teams to optimize inference for their specific hardware constraints. With support for frameworks like vLLM-Omni and SGLang, engineers can implement sequence parallelism to reduce latency on multi-GPU setups. This level of control is essential for scaling services to meet user demand while managing computational costs. The detailed deployment recipes provided by MiniMax lower the barrier to entry, enabling faster integration into existing product stacks.

What you can do

• Download the MiniMax-H3-Image checkpoint and text encoder from Hugging Face to inspect the model weights.
• Test local inference using SGLang or vLLM-Omni with the provided deployment scripts and configuration files.
• Experiment with the optional MiniMax-H3-Image-PE prompt expander to enhance short prompts before generation.
• Run the example Python scripts in the repository to compare local outputs with the hosted MiniMax API results.
• Configure sequence parallelism on multi-GPU systems to optimize latency for real-time editing applications.
• Review the MiniMax-H3-Image Community License Agreement to ensure compliance with your intended use case.