Pruna AI has released P-H3-Image, an open-weights text-to-image and editing model derived from MiniMax-H3. The release packages the model as a Diffusers ModularPipeline, generates images at up to 8 megapixels in eight steps, and accepts as many as eight reference images for compositional editing.

Each denoising step uses one model evaluation because P-H3-Image runs without classifier-free guidance. CFG normally combines conditional and unconditional predictions, roughly doubling the network work per step. Removing that overhead makes the eight-step pipeline materially faster, although its memory requirements remain substantial.

Memory determines the serving path

The bf16 package occupies about 100 GB and contains four main components:

• A 33.1-billion-parameter image transformer
• A 4-billion-parameter text encoder
• A variational autoencoder, or VAE, that converts between pixels and latent representations
• A Qwen3.5-9B-based prompt upsampler

Pruna reports the following performance for 8 MP generation on an Nvidia B200:

Direct bf16 inference therefore requires an H200, B200, or another accelerator with enough memory. Diffusers auto offloading lowers the reported peak to about 76 GB at 8 MP by moving components onto the GPU when needed, allowing the pipeline to run on a single 80 GB H100. Offloading also adds component transfers and host-memory use, so latency will differ from the direct B200 result.

Pruna also publishes quantized ComfyUI weights in int8 and 4-bit formats. Those variants use roughly 21 GB of VRAM, making local inference possible on smaller GPUs. The release does not provide a directly comparable timing result for that path.