Turning a large video model into a practical image generator is less about shrinking weights alone than removing work the image task does not need. This ComfyUI release reshapes MiniMax-H3 for still images, pairing an 8-step generation path with quantized components so high-resolution generation can run on substantially smaller hardware.
AI Image2026
P-H3-Image-Comfy
Generates high-resolution images and edits multiple references in ComfyUI in 8 steps. Based on MiniMax-H3, it offers int8 and w4a8 variants, with the smallest setup using about 21 GB of peak VRAM at 8 MP.
Introduction
Information
- Websitehuggingface.co
- OrganizationsPruna AI
- Published date2026/11/03
More Items
AI Model2026
Turns text descriptions into Three.js scenes that can be previewed and exported as GLB, targeting low-poly game assets and procedural environments. Fine-tuned from Qwen3.5-4B; some outputs may need retries or code fixes.
AI Model2026
Quantizes DeepSeek-V4.1-Flash-1017’s routed MoE experts and activations to NVFP4 for inference deployment. It supports multimodal input and up to one million tokens of context, but requires NVIDIA Blackwell GPUs and compatible runtimes.
AI Model2026
Compresses the model by pruning half its experts and quantizing the retained weights, targeting code, agentic workflows, long context, and vision. It retains 94.8%–98.7% of the base model’s scores on two coding benchmarks, at the cost of broader capabilities.