MiniMax-H3-Image FP8 fits a single 32GB AMD card: 4K images in 33s
evp-cloud · reddit · 2026-11-12
MiniMax-H3-Image's official BF16 weights need 120GB of VRAM, but the Paiton team built an FP8 version with their own runtime that runs on a single 32GB AMD Radeon AI PRO R9700:
- Speed: 2048×2048 in 6.9s; 4096×4096 in 33.4s; 4K edit with one reference image 39.9s, three references 88.5s.
- FP8 quality: over 30 prompts with identical starting noise, prompt adherence was essentially unchanged (CLIP 32.18 vs 32.31); differences were mostly composition, not quality — with only 6 steps, small numeric changes steer the sampler differently.
- Weak spots: small body text (newspaper columns become fake writing), close-up hands, clocks showing specific times.
- Deployment: Docker container with HTTP API and CLI; requires an R9700, 64GB RAM, 48GB disk. The FP8 weights only load in their runtime — no ComfyUI or diffusers support yet.
- License note: the H3 Community License does not cover the EU, UK, South Korea, or the US; hosted services need MiniMax's written authorization.
More from Infra
- Musk reportedly betting on Terafab: one plant producing a terawatt of silicon chips per year — XFreeze · 2026-11-12
- Precision on Demand: Escalating Weight Bits at Uncertain Decoding Steps — Mohammed Nayeem Khan · 2026-11-12
- Energy Tech Draws Nearly $200B in 2025 as Data-Center Loads Test Grid Interconnection Queues — ingliguori · 2026-11-12
- vLLM-Omni brings Day-0 support for MiniMax H3-Image: native 4K in 0.87s — vllm_project · 2026-11-12
- Anthropic boots Claude on AMD MI355X in ~3 days, fueling claims the CUDA moat is eroding — udmrzn · 2026-11-12
- Open Field Guide to Inference Engineering: KV Cache math to vLLM/SGLang production checklist — udmrzn · 2026-11-12