Chinese AI company MiniMax released the weights for its image generation and editing model, MiniMax H3-Image, in November 2026. The model combines text-to-image generation with reference-based generation and editing in a single model, and it can take up to 16 reference images. Its most notable feature is that it generates high-resolution images of roughly 16 megapixels directly, without an upscaling step. In addition to the API, it is available through MiniMax Code and Hailuo AI.
MiniMax says the model can generate a 4096×4096-pixel image in "under 10 seconds." However, an official comparison chart published in the same announcement shows 4K generation times that exceed 10 seconds depending on the task and mode. It is not clear whether the two figures were measured under the same conditions, so anyone judging real-world speed in a production environment will need to look closely at what was measured and how.
A unified model that generates and edits roughly 16-megapixel images
According to the model card published on Hugging Face, H3-Image supports text-to-image generation, new generation based on reference images, and editing of existing images. These capabilities are combined in a single checkpoint, so users do not need to switch between different models for different tasks. The figure of 16 refers to the number of reference images that can be input, not the number of images that can be generated at once.
Resolution is another key feature. H3-Image generates images of roughly 16 megapixels directly, rather than relying on an upscaler to enlarge them afterward. Official materials give examples of 4096×4096 pixels for a square image, 5472×3072 for 16:9, and 6272×2688 for 21:9. The supported range is about 4 to 20 megapixels, and the model card says results tend to be best at around 16 megapixels.
The model is built on MiniMax H3. The Diffusion Transformer (DiT) used for image generation has 33.1 billion parameters, derived from H3 with the audio-related layers removed, and it uses H3's text encoder as is. Note that the 33.1 billion figure is the DiT's parameter count, not the total for the whole system including the text encoder and other components.
Several measures were also taken to improve speed. According to MiniMax's official announcement, distillation reduced the number of sampling steps for image generation to six, and this was combined with improved training data, reinforcement learning, prompt expansion, and an optimized inference environment. However, MiniMax has not disclosed how much each technique contributed to the speedup. It has also not published run times on GPUs that users own themselves, so the official service and a local setup will not necessarily deliver the same speed.
The "under 10 seconds" claim versus the official benchmark
In its official announcement, MiniMax says it can generate a 4096×4096-pixel image in under 10 seconds. But in the official generation-time comparison chart on the same page, 4K generation times differ by task and mode. The table below extracts the H3-Image figures from the chart. Units are seconds, and P50 refers to the median processing time.
| Task | Standard | Highspeed |
|---|---|---|
| Text-to-image generation | 16.0 s | 10.3 s |
| Editing with 1 reference image | 18.2 s | 12.5 s |
| Editing with 3 reference images | 21.4 s | 14.4 s |
For 4K text-to-image generation, the median is 16.0 seconds in Standard mode and 10.3 seconds even in Highspeed mode. Editing with one reference image takes 18.2 and 12.5 seconds respectively, rising to 21.4 and 14.4 seconds with three reference images. At least within the official chart, editing with reference images takes longer than generating from text alone.
It is unclear, however, whether the "under 10 seconds" in the announcement text and the medians in the chart were measured under the same conditions. The difference in figures alone therefore does not show that either one is wrong. P50 is also a statistic indicating that half of all runs finish within that time, so generation will not always be that fast. These are figures published by MiniMax and have not been independently verified by third parties.
A footnote on the chart states that the second-round P50 was used for H3-Image's measurements, while original data was used for the other models. It is therefore not appropriate to read the chart as a speed ranking of all models measured in the same environment under the same conditions. The GPUs used, the content of the prompts, and whether pre- and post-processing were included in the measured time are not specified. Generation time when the maximum of 16 reference images is input also cannot be inferred from the published figures.
MiniMax also describes the model's image quality and instruction following as state of the art, but independent verification is limited. One notable gap is that no benchmark figures comparing instruction-following performance against other models are provided. Generation speed and output quality need to be evaluated separately, and the official announcement alone is not enough to judge overall merit.
For local use, converting prompts to English matters
When running the released model locally, users should be aware that input handling differs from the official API. The Hugging Face model card says H3-Image is designed on the assumption of "a detailed English description of the entire image." In other words, the processing flow differs between the official service, where short instructions can be entered in Japanese, and a local model that receives the same text directly.
The MiniMax API converts short instructions entered in any language into a detailed description of the whole image using a mechanism called "H3-Image-Context-IR" before generation. For local use, an optional prompt-expansion model called "PE" is also provided. However, there is no guarantee that using PE yields exactly the same preprocessing or output as the API. This difference matters when comparing image quality and instruction following between the API and a local environment.
When running locally with SGLang, the input prompt is passed to the model as is, and anything beyond 4,096 tokens is truncated. References such as "Picture 1" used in image editing correspond to the order in which reference images were uploaded. When handling multiple images, the image order must match the references in the text.
There are also constraints on inference parameters. num_inference_steps must be omitted or set to 6, and guidance_scale must be omitted or set to 1; any other value returns a 400 error. negative_prompt is not supported. The step-count tuning and negative-prompt output control common in other image generation models cannot be applied to H3-Image as is.
Preparing the runtime environment is also a considerable burden. The main model files total 68.65 GB, the text encoder, obtained from a separate repository, is 66.7 GB, and the optional PE prompt-expansion model is 17.9 GB. These are storage sizes, however, and do not directly indicate the GPU memory (VRAM) required, which varies with the inference framework, precision, and model loading method.
The model card introduces SGLang, vLLM-Omni, diffusers, and ComfyUI as execution environments. For vLLM-Omni, support for NVIDIA and AMD GPUs is described, with the H200, RTX PRO 6000 Blackwell, and MI350X given as examples. These notes do not guarantee that the model will run comfortably on typical consumer GPUs.
To compare API and local performance, the prompt, output resolution, number of reference images, and generation mode must be matched. In particular, whether preprocessing such as prompt expansion is counted in the measured time changes how long users actually wait. Additionally, the ability to convert Japanese instructions into detailed English descriptions is a separate matter from the ability to render correct Japanese text inside generated images, and the two should not be conflated.
Usable in Japan, but commercial offerings come with conditions
The release of H3-Image's weights does not mean the model can be used freely regardless of region or purpose. The license published on Hugging Face grants a free, limited usage license for all regions except the EU, the UK, South Korea, and the US. Japan is not among the excluded regions, so the model can be used there provided the license conditions are met.
Commercial use, however, carries additional conditions. If the user's or its affiliates' annualized revenue exceeds US$1 million, or if the user provides a hosted service to third parties, prior written permission from MiniMax is required. These are independent conditions: even with revenue of $1 million or less, prior permission is needed to offer the model as an API or web service used by third parties.
Commercial user interfaces must also display the model name. The license conditions to check therefore differ between using the model for in-house production work and embedding it in a service where outside users can generate images. The release of the weights should not be read as unconditional permission to build the model into commercial services.
Several clauses also need checking regarding rights to generated images. The license states that MiniMax claims no rights to generated outputs. At the same time, using, copying, modifying, distributing, or displaying the model or its outputs outside the licensed regions is not permitted. It is also prohibited to use the model or its outputs to improve the performance of other AI models, with the exception of H3-Image itself and its derivatives.
These are license terms applying to the released weights, and the same terms do not necessarily apply to the API that MiniMax operates. Anyone using the API should separately check the terms of service that apply to it. For those considering offering services to third parties or operating across multiple regions in particular, contractual restrictions may affect how the model can be deployed.
With direct generation at roughly 16 megapixels and support for up to 16 reference images, H3-Image broadens the options for high-resolution image production and multi-image editing. But judging its practicality requires looking beyond resolution to task-specific generation times, prompt preprocessing, GPU requirements, and license terms. As independent testing in local environments progresses, it should become possible to compare more concretely whether the official API or the released model is better suited to a given use.