P-H3-Image
Pruna's 8-step image and editing model, distilled from MiniMax-H3, with 20 GB 4-bit ComfyUI files and territory limits in the license.
In short. P-H3-Image is an AI image model you can download. It makes very large pictures, 8 megapixels or more, in only 8 steps, and it can edit a picture or combine up to 8 pictures into one scene. It was built by Pruna from MiniMax's video model H3, and the smallest version fits on a 24 GB graphics card. The license excludes the European Union, the United Kingdom, South Korea and the United States.
Summary
Pruna AI, a Munich-based company known for its model-compression library, released P-H3-Image on 3 November 2026. Its repositories first appeared on Hugging Face on 27 October. It is a text-to-image and image-editing model built from MiniMax's H3 video model, on the idea that a video model has already learned how objects look from different angles and how light falls on them. Pruna says it reworked every part of H3: the text encoder was distilled from 32B to 4B parameters (9 GB instead of 67 GB), sampling was distilled to 8 steps without classifier-free guidance, the transformer was slimmed by replacing its timestep block, and a new decoder turns one latent frame into a sharp image. A prompt upsampler based on Qwen3.5-9B is included and turns a short request into the detailed prompt the model renders best. It makes images from 2 to 16 megapixels, with 8 MP recommended, and edits with several input images, up to 8 on the model card and 9 in the launch post. The launch post says the transformer fell from 33B to 20B parameters, while the diffusers card still lists a 33.1B transformer; the ComfyUI files carry a 40.5 GB BF16 transformer, which fits 20B.
Pruna's own numbers are about speed and memory. It reports 2.6 seconds for a 4 MP image on one B200 against 12.4 seconds for MiniMax-H3, and 5.8 seconds against 30.6 seconds at 8 MP, where the diffusers card lists about 9 seconds for 8 MP plus 6 to 11 seconds for the prompt upsampler. It says that for most prompts the 8-step model matches the original 50-step one, and it publishes no benchmark score, so quality is a matter of looking at the images. Pruna calls it strongest at documentary and editorial photography, portraits, film stills, anime frames and graphic designs with short quoted text. Small or dense text is unreliable, hands and small background figures can break, results below 8 MP show more artifacts, and the model is sensitive to prompting, so hype words make images worse.
The ComfyUI repository holds bf16, int8 and w4a8 files. The bf16 set is a 55 GB download and peaks at 56 GB of VRAM at 8 MP, the int8 set is 30 GB and 31 GB, and the w4a8 set, which uses 4-bit weights with 8-bit activations, is 20 GB and about 21 GB, with a separate prompt upsampler of 19.3 GB in bf16, 9.7 GB in int8 or 6.7 GB in w4a8. Pruna says the int8 files stay very close to bf16 and the w4a8 files keep the quality but change more of the details. The diffusers pipeline is about 100 GB in BF16, needs about 101 GB of GPU memory with the upsampler or 83 GB without, and runs on one 80 GB GPU with offloading. In the 30 days to 11 November 2026, Hugging Face counted about 899,000 downloads of the ComfyUI repository and 1,400 of the diffusers one. The license is the MiniMax H3 Community License, which excludes the European Union, the United Kingdom, South Korea and the United States unless MiniMax gives you a separate license, and products must show "Powered by MiniMax H3". The Qwen3-VL-4B parts of the encoder and the Qwen3.5-9B upsampler are under Apache 2.0. Pruna also publishes P-Q-Image-2.1, LoRA adapters that let Qwen-Image-2.1 render in 6 steps.
Good for
- High-resolution photos, editorial and film looks, and graphic designs with a few quoted words
- Image editing and compositing from up to 8 reference pictures, from a one-line request
- ComfyUI users with a 24 GB graphics card, using the 4-bit set that peaks at about 21 GB
- Teams outside the excluded regions that want a faster, smaller route to the H3 look
Watch out for
- The license does not cover the European Union, the United Kingdom, South Korea or the United States, even though Pruna is based in Germany
- Commercial products must show "Powered by MiniMax H3", and the license has the other H3 conditions, so read it before shipping
- Pruna publishes speed and memory figures but no benchmark scores, and the claim that 8 steps match 50 is its own
- Small or dense text, hands and background figures can look wrong, and results get worse below 8 MP
- The launch post and the model card disagree on the transformer size and the maximum number of edit images
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of P-H3-Image. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| ComfyUI bf16 set BF16 | 20B | 55 GB | 56 GB | Workstation |
| ComfyUI int8 set INT8 | 20B | 30 GB | 31 GB | Gaming GPU |
| ComfyUI w4a8 set INT4 | 20B | 20 GB | 21 GB | Gaming GPU |
| Diffusers pipeline (BF16) BF16 | 33.1B | 100 GB | 76 GB | Workstation |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 21 GB of GPU memory.
The license in plain words
A community license that lets you use, modify and distribute MiniMax H3, but only in the "Applicable Territory": the whole world except the European Union, the United Kingdom, South Korea and the United States. Commercial products earning over US$20 million a year need MiniMax's written authorization, every commercial product must show "MiniMax H3" in its interface, and you may not use the model or its outputs to train other AI models. Get legal advice, especially if you or your users are in an excluded region.
- Attribution or model name must be shown
- Free commercial use only below a revenue cap
- Not licensed in some regions
- Acceptable-use policy applies
- Outputs must be filtered or reviewed, and AI generation disclosed
More about MiniMax H3 Community
See also
Frequently asked questions
- What is P-H3-Image?
- P-H3-Image is an AI image model you can download. It makes very large pictures, 8 megapixels or more, in only 8 steps, and it can edit a picture or combine up to 8 pictures into one scene. It was built by Pruna from MiniMax's video model H3, and the smallest version fits on a 24 GB graphics card. The license excludes the European Union, the United Kingdom, South Korea and the United States.
- Can I use P-H3-Image commercially?
- Yes, with conditions. P-H3-Image is released under the MiniMax H3 Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Not licensed in some regions; Acceptable-use policy applies; Outputs must be filtered or reviewed, and AI generation disclosed. Read the license text before you ship a product.
- Is P-H3-Image open source?
- P-H3-Image is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax H3 Community adds conditions that the Open Source Initiative definition does not allow.
- Can I run P-H3-Image on my own computer?
- Yes, on a good graphics card. The smallest variant, ComfyUI w4a8 set, needs about 21 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was P-H3-Image released?
- P-H3-Image was released by Pruna AI on November 3, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- P-H3-Image: Turning a SOTA Video Model into an Efficient Open Image Model (Pruna AI, 3 Nov 2026) pruna.ai
- P-H3-Image model card (Hugging Face) huggingface.co
- P-H3-Image-Comfy model card (Hugging Face) huggingface.co
- Pruna AI legal notice (company registration in Munich) pruna.ai
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| MiniMax H3-Image MiniMax | 33.1B | – | MiniMax H3-Image Community (Conditional) | 155 GB Big-memory desktop | |
| MiniMax H3 MiniMax | 33.1B | – | MiniMax H3 Community (Conditional) | 166 GB Big-memory desktop | |
| HyperFlow H3 Flash Video Rebirth | 13.7B – 13.9B | – | MiniMax H3 Community (Conditional) | 18 GB Gaming GPU | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone |