MiniMax open-sources a 27B prompt rewriter for its H3 image model

MiniMax has published MiniMax-H3-Image-PE on Hugging Face, a 27B-parameter model that rewrites a short image request into a detailed description and suggests an aspect ratio for MiniMax-H3-Image. The weights are FP4 and fine-tuned from Qwen3.8-27B.

What the release contains and how to run it

The repository covers two workflows. For text-to-image prompt expansion, MiniMax ships the weights, a task prompt, a system prompt and a client example. For image editing, it ships only a task prompt and a message-assembly example meant for a vision-language model of the user’s choice. MiniMax says editing expansion is still under development, so there is no editing model yet.

The company recommends serving the model with SGLang 0.5.23 or newer and XGrammar 0.2.7 or newer. XGrammar forces the answer into a JSON schema, and a Qwen3-style reasoning parser splits the model’s reasoning from the final reply. The text-to-image output has two fields, expanded_caption and aspect_ratio. In MiniMax’s example, the request for an orange tabby cat on a sunlit windowsill returns a one-sentence caption and a 3:2 ratio.

The default request settings are fixed in a file called prompt/t2i/request_params.json. They include temperature 1.0, top_p 0.95, top_k 20, a 4096-token budget that covers both thinking and the answer, and reasoning_effort set to xhigh. The chat template also accepts medium and low effort levels. The included script scripts/expand_prompt.py sends the same request and validates the response.

The ratio logic is simple. If the user’s aspect-ratio setting is auto, the pipeline should use the model’s recommendation. If the user picks a ratio, that choice wins at generation time. For editing, the prompt takes the instruction, at least one reference image with its source ratio, and a target ratio that is either adaptive or an explicit W:H value. The returned field there is target_aspect_ratio, and an explicit ratio is preserved exactly.

MiniMax also points to hosted options: the H3-Image models are reachable through platform.minimax.io globally and platform.minimax.cn in China, with documentation for text-to-image, editing and prompt expansion.

Licensing leaves commercial use to a form

The model is not under a standard open license. MiniMax releases it under the MiniMax-H3-Image Community License Agreement, while noting that its base, Qwen3.8-27B, is Apache 2.0. Section IV of MiniMax’s license describes separate written authorization, which is requested at platform.minimax.io/license. The page does not say which uses need that approval, so teams planning commercial deployment should read the license text before building on it.

A small companion to a very large image model

The release is easier to judge next to its parent. On the MiniMax-H3-Image page, published on Nov. 3, the company describes a 33.1B-parameter diffusion transformer that generates roughly 16-megapixel images directly, from 4096×4096 up to 6272×2688, in six sampling steps without classifier-free guidance. It accepts up to 16 reference images for editing, and its BF16 checkpoint is listed at 68.65 GB. The same page lists the PE model at 17.9 GB in NVFP4 as an optional add-on, and says short prompts in any language can instead go through MiniMax’s hosted expansion endpoint.

That context explains the design. The image model is tuned for detailed English prompts, so a person typing a few words gets weaker control than a person writing a paragraph. A prompt expander closes that gap, and shipping it as weights lets self-hosters avoid sending prompts to a hosted API. The cost is a second model: roughly 18 GB of extra weights and a reasoning step with a 4096-token budget before each image, according to the two model pages.

The choice of base is also worth noting. Qwen3.8-27B is the same-size open model that rival releases measure themselves against, as in our report on Nanbeige4.3-33B, and our explainer on Alibaba’s Qwen family covers where that lineage comes from. Here a Qwen-derived text model sits in front of a diffusion model whose text encoder is also Qwen-based, per MiniMax’s H3-Image page, so the whole chain leans on Alibaba’s open releases.

What MiniMax has not shown

The announcement carries no benchmark results, no human preference tests and no comparison of images generated with and without expansion. The H3-Image page makes a qualitative claim of the strongest instruction following among open-source models but gives no scores for it, and nothing in either repository measures how much the PE step contributes. Both claims are MiniMax’s own and unverified. The editing workflow cannot be tried as a model at all, since only its prompt is released. At the time of the page capture the repository showed zero downloads and two likes, so community testing has barely started.

For readers who run image pipelines, the practical question is whether an extra 27B reasoning pass is worth the latency for each request. The release gives the tools to find out, but not the answer.