model release

MiniMax releases MiniMax-H3-Image-PE, a prompt-expansion model fine-tuned from Qwen3.8-27B in FP4

TL;DR

MiniMax has released MiniMax-H3-Image-PE on Hugging Face, a prompt-expansion model for MiniMax-H3-Image. The weights are fine-tuned from Qwen3.8-27B in FP4 and turn a short request into a detailed image description plus an aspect-ratio recommendation.

3 min read
0

MiniMax has released MiniMax-H3-Image-PE, a prompt-expansion (PE) model for its MiniMax-H3-Image generator, with open weights on Hugging Face. The weights are fine-tuned from Qwen3.8-27B and shipped in FP4. The model turns a short request into a detailed description of the intended image and recommends an aspect ratio for generation.

What was released

The repository covers two workflows:

  • Text-to-image (T2I) PE: takes a text instruction. Released assets are the MiniMax-H3-Image-PE weights, task and system prompts, and a client example.
  • Image editing PE: takes an editing instruction and reference images. Only a task prompt and a message-assembly example are released, for a vision-language model of your choice. MiniMax states that image editing PE is still under development and this release provides its prompt only.

MiniMax also lists API access to MiniMax-H3-Image through platform.minimax.io (global) and platform.minimax.cn (China), with documentation for text-to-image, image editing and prompt expansion. API pricing is not yet disclosed.

Serving and request settings

MiniMax recommends SGLang (sglang>=0.5.23) with xgrammar>=0.2.7. SGLang reads the FP4 configuration from the checkpoint. XGrammar constrains the answer to a JSON schema, and the qwen3 reasoning parser separates reasoning from the final answer.

Default request parameters in the repository:

Parameter Value
temperature / top_p / top_k 1.0 / 0.95 / 20
min_p 0.0
presence_penalty / repetition_penalty 0.0 / 1.0
max_tokens 4096, including thinking and the final answer
reasoning_effort xhigh
enable_thinking / preserve_thinking true / true

The T2I output is a JSON object with expanded_caption and aspect_ratio. The example in the repository expands "An orange tabby cat sleeping on a sunlit windowsill." into a one-sentence photographic description with a 3:2 ratio. If the aspect-ratio selection is auto, the returned ratio should be used for image generation. If the user selects an explicit ratio, that ratio takes precedence at the generation stage.

Image editing prompt

The editing workflow requires a vision-language model that supports interleaved text and images. Inputs are the editing instruction, at least one reference image with its source ratio as a reduced W:H string (or None if unknown), and a target ratio that is either adaptive or an explicit positive-integer W:H string. The fixed task prompt forms a shared prefix across requests, which allows reuse by services that support prefix caching; cache behavior depends on the provider.

The output contains expanded_caption and target_aspect_ratio. The caption identifies source subjects, states what is preserved or changed, and describes new content. Descriptive prose is English, while literal text keeps the language the user requests. An explicit ratio is preserved exactly, and with adaptive the model determines a numeric ratio.

Local use with Diffusers

MiniMax-H3-Image-PE can run locally in Transformers through the diffusers/minimax-h3-image-pe-blocks Modular Diffusers blocks, which use the released prompts and request parameters for both workflows. The FP4 weights stay in 4-bit and use about 19 GiB of GPU memory. GPUs with native FP4 support run the NVFP4 weights directly, and other GPUs use bitsandbytes FP4.

What this means

This is a companion model rather than an image generator: it sits in front of MiniMax-H3-Image and handles prompt rewriting and aspect-ratio selection. Releasing the weights, prompts and request parameters lets developers run that step on their own hardware instead of through the hosted API. The roughly 19 GiB FP4 footprint, as stated by MiniMax, puts local deployment within reach of a single GPU. Because image editing PE ships as a prompt only, teams that need editing support must supply their own vision-language model for now.

Related Articles

model release

Shanghai AI Laboratory releases Sage S2.1 Flash, a 321B-parameter scientific multimodal model under MIT license

Shanghai Artificial Intelligence Laboratory has published Sage S2.1 Flash, a 321B-parameter scientific multimodal model with MIT-licensed weights and an FP8 variant. According to the lab, it placed first on four of 15 charted benchmarks and second on seven, including 90.3 on MLE-bench Lite.

model release

Meituan releases LongCat-Video-Avatar-1.6, open-weight avatar model for 10-minute-plus videos under MIT license

Meituan has released LongCat-Video-Avatar-1.6, an open-source audio-driven human video generation model with weights under the MIT License. The company claims it sustains videos longer than 10 minutes and ranks second among six systems in a human preference study of 33,051 valid votes.

model release

Liquid AI releases d1.2 and d1.2-ultra, decision models that return probabilities with zero output tokens

Liquid AI released d1.2 and d1.2-ultra, decision models that read text, images and audio and return calibrated probabilities with zero output tokens. d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1.

model release

Huawei Cloud releases CloudRobo-WM1.2 world action model, claims first place on RoboTwin 2.0 and LIBERO-Pro 2.0

Huawei Cloud has released CloudRobo-WM1.2, an embodied world action model, and says it will be open source. The company claims first place on the RoboTwin 2.0 and LIBERO-Pro 2.0 robot manipulation benchmarks, with scores of 88.98% and 81.7%.

Comments

Loading...