A repository release
What exists is the repository itself: created on November 8, first content commit on November 10 at 14:38 UTC, and a second commit 38 minutes later adding a Diffusers section via a merged pull request. The worker found no separate launch post for the expander in MiniMax's news index checked November 11, which lists H3, M3.1 Flash, Speech 2.8 and Music 3.0, and nothing about images.
At the desk's November 11 check, around 05:05 UTC, Hugging Face displayed zero downloads and 36 likes for the expander; the parent model displayed 667 downloads. Displayed metrics on a repository this young say little either way. They do mean every quality statement about this model currently traces back to its own card: no benchmark, Elo or preference score for the expander appears on either model card, and the sources reviewed supply no independent evaluation of it.
What the expander is configured to produce
The parent model explains why a companion exists. H3-Image is distilled to exactly six sampling steps, with no classifier-free guidance, and its served API rejects other step counts, other guidance values and negative prompts outright. References and dimensions still matter, but the prompt carries unusual weight in this stack: the parent card says it expects a detailed English description of the whole image, and the local serving path truncates input at 4,096 tokens.
MiniMax's answer is to move the interpretation upstream, into a model that can reason before it writes. The supplied request configuration enables thinking, sets reasoning effort to "xhigh", preserves the thinking, and caps output at 4,096 tokens, with only the final JSON constrained: exactly two fields, an expanded caption and an aspect ratio from a fixed list of ten. The aspect-ratio decision is specified just as tightly in the task prompt. It sets a precedence order — an explicit final canvas first, then the requested deliverable, then the subject's composition — and separates a depicted scene from a communication artifact, so a request for a results slide is built as an information layout rather than photographed as a meeting room. Anyone who has watched an image model do the second of those will recognize the problem being solved.
The parent card also links to documentation for a hosted equivalent, H3-Image-Context-IR. Comparing local expansion with that service requires its current pricing, behavior and resource costs, none of which the cards state.
Editing prompts are included; editing weights are not
The repository describes two workflows, and they are not the same offer. Text-to-image expansion is documented for execution: weights, system and task prompts, parameters and a client script. Image editing expansion ships a task prompt and a message-assembly example, to be run through a vision-language model of your own choosing. The card says so directly: "Image editing PE is still under development; this release provides its prompt only."
Read the repository name with that split in mind. A team that downloads 17.92 GB expecting help with reference-based edits — the workflow H3-Image itself promotes with up to 16 reference images — gets a prompt instead. The prompt is specific: it requires each reference image's reduced width-to-height ratio as an input and tells the model not to estimate that ratio from a preview. But the sources reviewed supply no evaluation of how well any specified vision-language model follows it.
The license travels with it
The expander is governed by the MiniMax-H3-Image Community License Agreement, and its repository carries that license file, dated November 10, 2026. That imports the terms covered in our H3-Image article: the grant applies only in an Applicable Territory defined as the world excluding the European Union, the United Kingdom, South Korea and the United States, and Section IV requires prior written authorization if annualized revenue exceeds US$1 million, or if you provide a hosted service to third parties at all, with no revenue floor on the second condition. Section IV also requires commercial products using the works to display "MiniMax-H3-Image" prominently on the interface. A prompt writer sounds like the harmless part of a stack; under this license it is subject to the same stated conditions as the generator it serves.
What adopting it actually commits you to
Three commitments, stated plainly. You run two models instead of one: the expander's 17.92 GB of weights and roughly 19 GiB of GPU memory while it runs, on top of an H3-Image stack whose downloads already total roughly 135 GB once its borrowed text encoder is counted. You accept the supplied configuration's reasoning setup — thinking on, effort at xhigh — as the starting point for every expansion, which is a latency decision an interactive product has to hide or pre-compute. And you accept that there is no published measurement of whether the expanded captions produce better images than a shorter prompt or a smaller expander. Until someone runs that comparison, the case for this model rests on its design logic, not on results.
Sources
- [1] Hugging Face model card and repository, MiniMaxAI/MiniMax-H3-Image-PE (workflows, request parameters, serving, license statement; repository metadata and files read first-hand by the ALFW desk, November 11, 2026)Read source
- [2] Hugging Face model card, MiniMaxAI/MiniMax-H3-Image (parent model expectations, six-step sampler, Context-IR documentation link)Read source
- [3] MiniMax-H3-Image Community License Agreement (Applicable Territory; Section IV authorization and attribution)Read source
- [4] OrcaRouter, "MiniMax-H3-Image-PE: a 27B prompt expander, unannounced" (repository timestamps and file inventory, November 11, 2026)Read source