Running MiniMax-H3 on AMD Strix Halo: 5s clip in 9.5 min, but times vary
danielcar · reddit · 2026-08-14
Developer danielcar ran MiniMax-H3 (video+audio) locally on AMD Strix Halo with 128GB unified memory and no discrete GPU, using ROCm 7.14, ComfyUI, int8 pruned transformer, and 4-step turbo LoRA. A 5-second 896×512 clip took 9.5 minutes, but subsequent generations took 10 minutes and 1 hour respectively. The author notes weights are community pruned conversions, so quality isn't representative of official H3; this was a speed/setup test. Scripts and full writeup shared.
More from Embodied
- RLBotics: Lightweight GPU-Accelerated RL Framework for Isaac Lab — rsasaki0109 · 2026-08-14
- Cloud Workstations Reshape Embodied AI R&D Infrastructure — 量子位 · 2026-08-14
- NVIDIA's ProtoMotions: GPU-Accelerated Simulation Framework for Humanoids — carlosdponx · 2026-08-14
- AI-generated circuit fails: Developer wastes an hour due to AI's design flaw — debreuil · 2026-08-14
- Tesla AI Captures Solar Eclipse via Autonomous Driving Vision System — bendomingueztv · 2026-08-14
- Stop the Humanoid Hype: Purpose-Built Arms Outperform Figure in Logistics — freelerobot · 2026-08-14