Running MiniMax H3 on an 8 GB laptop GPU — making it work, making it stable, and publishing every number measured along the way.
A working 8 GB laptop configuration for H3, published in full https://github.com/FlowForgeLabAi/MiniMax-H3-8GB-Workflows
Three tiers: 20-step, 8-step, 4-step. A 10 s 768×1024 clip takes 14.3 min at 8 steps and 33.8–43.5 min at 20 steps.
Each trap on that path written up as a reproducible case https://github.com/FlowForgeLabAi/-h3-8gb-traps
Six measured traps, each with its reproduction. The central finding: what stops you is not the 8 GB of VRAM, it is the 16 GB of system RAM.
Raw telemetry, published https://huggingface.co/datasets/FlowForgeLabAi/minimax-h3-8gb-bench
18 timed runs plus 534 samples of GPU monitoring data.
A self-check page, indexed on the H3 model page https://huggingface.co/spaces/FlowForgeLabAi/minimax-h3-8gb-advisor
The six traps, a power-draw self-test, and a wall-clock predictor, T ≈ 102 + 34.1 × steps.
A ComfyUI bug report that an official fix PR cites Comfy-Org/ComfyUI#16544
SaveVideo fails to open libx264 when width or height is odd. The fix PR #16545 carries
Fixes #16544 in its description.
A technical exchange in the official H3 repository https://huggingface.co/MiniMaxAI/MiniMax-H3/discussions/104
On the wall that a "stream from disk to run H3 under 4 GB of VRAM" approach runs into, and how to tell whether you have hit it.
| Finding | Number |
|---|---|
| The bottleneck is system memory, not VRAM | Working set ≈ 34.5 GiB (19.5 GiB model + 15.0 GiB text encoder) against 15.26 GiB of RAM |
| GPU utilisation lies when it stalls | Healthy: 64–98 W. Stalled: flat 34–36 W while utilisation still reads 99% |
| Paging degrades throughput non-linearly | Same graph, same resolution: 0.9–6 s/block fresh, 212 s/block after ~1.7 h |
| Attention backends differ enormously at H3's real geometry | SDPA 143.1 ms / comfy-kitchen INT8 22.8 ms / SageAttention 29.0 ms |
| A LoRA applies to only part of a pruned model | 208 of 259 modules (80.3%); all 51 adaln_proj.linear skipped |
| The pinned-memory budget on Windows | MAX_PINNED_MEMORY = ram * 0.40 (≈6.1 GiB non-pageable) |
After publishing, I went back to the source and found I had quoted the wrong constant: I had
written ram * 0.90, which is the Linux branch. Windows takes ram * 0.40.
Before writing the correction I first confirmed which branch I was actually on — because the wrong branch gives 11.26 GiB and the right value is 6.10 GiB.
I corrected three published artefacts and kept the one copy that had the platform right all along. The full write-up is here: https://github.com/FlowForgeLabAi/-h3-8gb-traps/blob/main/verification-and-correction-2026-10-01.md
This section stays in on purpose. It is evidence of how I work, not a mark against it.
- The usable envelope for H3 on 8 GB: resolution, step count, reference images
- Turning the measurements above into a one-page self-check other people can run