MiniMax H3 Image's ~50GB stack runs on 8GB VRAM at 57s/image — how is there no OOM?
wreck_of_u · reddit · 2026-11-14
A user ran MiniMax H3 Image (int8-pruned, 50GB total stack) on an RTX 2070 8GB with 32GB DDR4 and a first-gen Ryzen 1700, doing i2i with two 1024x1024 reference photos at 57 seconds per image in ComfyUI — faster than their Krea2, Flux2 Klein, or Qwen 2.1 edit workflows. They're puzzled why it doesn't OOM, since shouldn't all models, including the massive 32B text encoder, load entirely into memory? The post is a question with no authoritative answer yet.
Related event: MiniMax H3 Image Runs on 8GB GPUs After Quantization(3 posts)→
More from Infra
- Marvell to show photonic fabric memory at SC26 with sub-350ns cross-rack latency — zephyr_z9 · 2026-11-14
- Japanese stock markets fall back to fax machines as Microsoft outage hits hour 12 — gleech · 2026-11-14
- Analyst weighs AI industry fallout after Nvidia scandal and Jensen Huang's resignation — chrismattmann · 2026-11-14
- Crypto proofs emerge as key tools for AI verification, safety and governance — DavideCrapis · 2026-11-14
- 'This is Enron 2.0': viral thread claims Nvidia finances its own demand amid $200B of idle GPUs — churchkey · 2026-11-14
- Megaspeed still listed as preferred Nvidia cloud partner on official site — bigdata · 2026-11-14