MiniMax H3 on 12-16GB GPUs: seven speed-up options compared

musetta_ai · reddit · 2026-10-26

The author compiled seven speed-up options for running MiniMax H3 video generation on 12-16GB cards, all from official model cards. Core insight: on small cards the bottleneck is usually weights or the text encoder streaming from system RAM, and turbo LoRAs cut steps, not per-step time (per ComfyUI docs).

Key points:

• Base files: bf16 66.28GB down to prunedw6a8 at 15.98GB; prefer int8convrot on cu130 PyTorch; CPU offload costs more than precision tradeoffs when VRAM is tight
• The text encoder is the hidden killer: official Qwen3-VL 32B nvfp4awq is just 15.69GB; the GGUF encoder (Q4KM, 19.8GB) with post-conditioning eviction turned a 90-min stall into a 12.3-min render on a 5090 — check this first if a big card is slow
• Turbo LoRA picks: Comfy-Org LightX2V (8 steps T2V/I2V, 4 R2V), Larry Turbo v4 (4-8 steps; over 8 gets over-sharp; smear 1.05-1.2), DMAD (4 steps, cfg 1.0, T2V only, 210s for 5s clip on 8GB)
• Two-stage route (tested on 16GB 3080 laptop): molbal merged turbo GGUF (Q40, 11.4GB), sample 2 of 4 steps at 0.2MP, latent-upscale to 1.0MP, finish last 2 steps
• FastH3 is not for small cards: 8-Step V2 needs the VSA backend tested on four B200s; QAD-v2 is Blackwell-only

Free extras: Sage Attention (2x per docs).

Original post →

More from Multimodal

• Infinite Recursion Vortex prompt makes images spiral into themselves — LudovicCreator · 2026-10-26
• Dreamina offers $0.29 first month for Seedance, sparking debate on AI video subscription value — AmbitiousMeat6860 · 2026-10-26
• MiniMax amplifies workflow claim: Opus 5.5 + H3 mass-produces viral Douyin MVs in minutes — Hailuo_AI · 2026-10-26
• Tencent Hunyuan open-sources Hy-SimulMT-30B: bidirectional CN-EN live speech translation with voice cloning — jacek2023 · 2026-10-26
• Midjourney sref combo --sref 1240652461 369650292 yields soft snowy illustration styles — sergeantsref · 2026-10-26
• Skip Figma for gradients — Midjourney generates stunning gradient backdrops — hewarsaber · 2026-10-26

Multimodal channel →