Running MiniMax-H3-Image locally: 16MP image in 6 steps, ~22s on a 4090
metover · reddit · 2026-11-12
A Reddit user shared a full ComfyUI setup for running MiniMax-H3-Image locally at 16.6MP (4160x4000): pruned bf16 DiT weights, Qwen3-VL-32B text encoder with the official VAE, euler/simple sampler with just 6 steps at CFG 1, single seed, no upscaling or inpainting, and hand-written 4-5 sentence prompts without a prompt expander. Speed: 20s/image on an RTX PRO 6000, 22s on a 4090 with int8 weights.
Related event: Users Test MiniMax H3 Image Model Locally with Strong Results(2 posts)→
More from Infra
- HP starts selling $132,000 GB300 desktop with 748GB memory for local AI — gnukeith · 2026-11-12
- Dev runs out of GitHub Actions minutes, shows off Mac Studio with unreleased M5 Ultra chip — ow · 2026-11-12
- NVIDIA + RadixArk boost SGLang Rubin performance up to 10x in under 5 days — ying11231 · 2026-11-12
- How Muse, Tailscale and MCP let a cloud AI agent securely reach a home network and drive a radio — Dave_Maynor · 2026-11-12
- AI capability gains per billion dollars falling 8x in 2 years — and ASI may never be general, argues ramez — burny_tech · 2026-11-12
- AWS ships official local emulator built by agent swarm: 29,684 commits in three weeks — Simon Willison · 2026-11-12