Testing Minimax H3 locally: runs 15s video gen on 12GB VRAM
ayakitodev · reddit · 2026-08-03
The author shares a complete workflow and parameter configurations for running the Minimax H3 image-to-video model locally in ComfyUI.
- Hardware & Usage: On a Windows 11 machine with 16GB VRAM and 96GB RAM, the generation peaked at around 12GB VRAM and 91GB system RAM.
- Key Parameters: Used minimaxh3ref2vabf16 model and qwen3vl32b text encoder with Euler beta57 sampler, taking 20 steps to generate a 960x564 resolution, 15-second video.
- Pitfall: Warns users not to accidentally feed the negative prompt generated by ChatGPT into the model.
Related event: MiniMax H3 Local Tests: Runs Video Generation on 12GB VRAM(2 posts)→
More from Infra
- NVIDIA B300 Specs Leaked via nvidia-smi: 283GB VRAM, 1100W TDP — Maximus-CZ · 2026-08-03
- Half a Million Sites Block AI Crawlers: How to Check Your Training Data Status — VeryWellVersed · 2026-08-03
- Quick Tip: Launch llama.cpp Router Mode Fast via Windows Search — Addyad · 2026-08-03
- MiniMax H3 Local Test: 182s for 5s Video on RTX 4090 Laptop — robomar_ai_art · 2026-08-03
- Cloudflare Kicks Off Agents Week: Envisioning an 'Agent Cloud' for Autonomous AI — threepointone · 2026-08-03
- Flipkart Founder: Civilizational Prosperity is Directly Proportional to Energy Consumption — NirantK · 2026-08-03