Is P-H3-Image state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
Pruna AI's open-weights text-to-image and image-editing model, rebuilt from the MiniMax-H3 video model: 8 sampling steps instead of 50, no classifier-free guidance, 2 to 16 MP output and edits with up to 8 input images. Weights on Hugging Face under the MiniMax H3 Community License, for Diffusers and ComfyUI.
The details so far
- Lab
- Pruna AI
- Announced
- Nov 3, 2026
- API
- Not yet · No hosted API stated in the launch materials
- Open weights
- Published · Hugging Face PrunaAI/P-H3-Image (Diffusers ModularPipeline, about 100 GB in bf16) and PrunaAI/P-H3-Image-Comfy (int8 and 4-bit files for ComfyUI); not gated; a Playground on Hugging Face Spaces
- Parameters
- 33.1B-parameter transformer per the card (66 GB in bf16), plus a 4B text encoder and a prompt upsampler based on Qwen3.5-9B; Pruna's blog describes the backbone as 20B after replacing H3's timestep block
- Type
- Text-to-image and image editing (1 to 8 input images); single images only, no video or audio
- Base model
- MiniMaxAI/MiniMax-H3 (a video model), fine-tuned and reworked: prompt upsampler, text encoder, diffusion schedule, transformer and decoder
- Sampling
- 8 steps (distilled from 50), no CFG or negative prompt
- Resolution
- 2 to 16 MP; the card recommends 8 MP for text-to-image and 4 or 8 MP for edits
- Text encoder
- 4B, distilled from H3's 32B encoder (9 GB vs 67 GB per the blog)
- Prompt upsampler
- Included and on by default, based on Qwen3.5-9B
- Hardware
- About 101 GB GPU memory at 8 MP in bf16 with the upsampler (83 GB without), or 80 GB with offloading; the 4-bit ComfyUI build runs in about 21 GB
- License
- MiniMax H3 Community License: excludes the EU, UK, Republic of Korea and US without a separate MiniMax license and requires "Powered by MiniMax H3"; Qwen-derived parts Apache 2.0, ComfyUI nodes MIT
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 5x faster than MiniMax-H3 with 8 steps instead of 50: a 4 MP image in 2.6 s vs 12.4 s on one B200 (5.8 s vs 30.6 s at 8 MP).
Lab claim: 7x smaller than H3 at its smallest (a 20 GB 4-bit download vs 143 GB), running 8 MP images in about 21 GB of VRAM.
Lab claim: for most prompts, the 8-step model matches the 50-step original.
Sources
LABP-H3-Image launch post (Pruna blog)pruna.ai/blog/oss-p-h3-image-launch · Nov 3, 2026LABPruna AI announcement on Xx.com/PrunaAI/status/2117663564487926043 · Nov 3, 2026REPOPrunaAI/P-H3-Image on Hugging Facehuggingface.co/PrunaAI/P-H3-ImageREPOPrunaAI/P-H3-Image-Comfy on Hugging Facehuggingface.co/PrunaAI/P-H3-Image-ComfyTracking continues until an independent board lists it.