Pruna AI Shrinks MiniMax-H3 From 143 GB Down to a Single GPU
Pruna AI turned MiniMax's 143 GB H3 video model into an open image generator and editor that renders 8 MP images in 8 steps and fits in 21 GB of VRAM.
- Pruna AI released P-Image-OSS, an open-weights image model rebuilt from the MiniMax-H3 video model.
- It makes 8 MP images in 8 steps without CFG, about 5x faster than H3's 50 steps.
- The smallest 4-bit set is a 20 GB download peaking at 21 GB VRAM, versus 143 GB for H3.
- Pruna distilled the text encoder from 32B to 4B, cut the transformer to 20B, and trained a new decoder.
- It edits with up to 8 reference images and runs natively in ComfyUI and Diffusers.
- The MiniMax license excludes the EU, UK, South Korea and US without a separate MiniMax license.
Pruna AI converted MiniMax-H3, a video model with 143 GB of weights, into an image generation and editing system. The resulting P-Image-OSS, initially announced as P-H3-Image, has recorded more than 1.2 million downloads through its ComfyUI repository on Hugging Face.
P-Image-OSS generates 8-megapixel images in eight denoising steps, the iterative passes that turn noise into an image. It runs without classifier-free guidance, a common technique that evaluates conditioned and unconditioned predictions at each step and roughly doubles denoiser work. The model can also edit with up to eight reference images. Its smallest package is a 20 GB download with a reported peak of about 21 GB of VRAM, placing it within the memory budget of a 24 GB graphics card.
Why a video backbone helps
Video models must preserve objects, lighting, materials and anatomy as scenes change across frames. That training gives them useful representations of three-dimensional shape, motion and viewpoint, which can transfer to still-image generation. Pruna previously tested the approach with Wan Image, built from Wan 2.1, before applying it to MiniMax-H3 after MiniMax released the weights.
H3’s original design created two obstacles for image generation: individual frames appeared soft, and the model required too much memory and compute for common local hardware. Pruna’s technical report describes changes to the prompt upsampler, text encoder, diffusion schedule, transformer and decoder.
Five changes behind the conversion
- Prompt upsampler. H3 responds strongly to prompt structure. Pruna trained a Qwen3.5-9B-based model to expand a short request into the detailed format H3 handles well. It processes the request before drafting the expanded prompt, uses greedy decoding for repeatable output and limits prompt length to reduce glossy or overworked images.
- Text encoder. H3 originally interpreted prompts with a 32-billion-parameter encoder. Pruna distilled that encoder into a 4-billion-parameter model by training the smaller network to reproduce the larger model’s representations. The reported footprint fell from 67 GB to 9 GB.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.