Community shrinks open-source MiniMax-H3-Image from 68GB to 6.4GB, 110k downloads
alexcovo_eth · x · 2026-11-13
MiniMax's newly open-sourced 33.1B image model MiniMax-H3-Image ships at 68.65GB — out of reach for most home PCs. Within days, the community produced 27 quantized versions with over 110k downloads.
Highlights from hands-on reports:
- Much lower VRAM barrier: Q4 runs at 10.78GB, extreme Q2K at just 6.4GB (even thin Macs); Unsloth FP8 (18.8GB) maximizes quality.
- Native 4K output: 4096×4096 (16MP) directly, no upscaler needed, including 21:9 cinematic ratios.
- 6-step fast generation: officially distilled to 6 sampling steps, no CFG required.
- Generation + editing: supports up to 16 reference images for character control; ComfyUI GGUF and Mac MLX weights ready.
Quantized collection is live on Hugging Face — a strong option for 8GB/12GB GPUs wanting high-res generation.
More from Multimodal
- Free Booru Tag Atlas for NoobAI-XL: 2,900+ tags in EN/RU/ZH — tony_cash_original · 2026-11-13
- SpAItial unveils Segmentation API that splits Gaussian splat worlds into labeled meshes — angelaqdai · 2026-11-13
- RTX 5070 Ti vs RTX 3090 in ComfyUI: 5070 Ti wins warm batches, 3090 wins cold loads — Numerous_Access_8008 · 2026-11-13
- Pruna AI open-sources object-removal LoRAs for Qwen-Image 2.1, 6 steps vs 40 — pbaylies · 2026-11-13
- Suno launches Bridge beta, bringing its music models directly into Ableton Live 12 — suno · 2026-11-13
- SpaceX lays off ~60 on Grok Imagine team as Musk shifts focus to AI coding and agents — GaryMarcus · 2026-11-13