Introduction

Most MiniMax H3 guides assume an NVIDIA card. If you have a Radeon, there's finally a measured, reproducible path.

minimax-h3-rdna2-windows went up on GitHub on Nov 1 (15:43 UTC), with mirrors on Hugging Face and ModelScope. It's an unofficial, MIT-licensed set of PowerShell scripts and engineering notes for running MiniMax H3, the 33B video-plus-audio diffusion model, on a single AMD RX 6900 XT (RDNA2, 16 GB) on Windows. It uses stable-diffusion.cpp with AMD's ROCm/HIP backend. The headline number: a 480p, 2.3-second clip went from 3,707 s to 183 s, a 20.3× speedup on the same card.

What's in the Box

  • gen.ps1: the main generator. It has four presets, picks the backend and budgets VRAM automatically.
  • serve.ps1: a local web UI at 127.0.0.1:1234.
  • extract_rocm.ps1: pulls the ROCm runtime out of AMD's HIP SDK installer without admin rights.
  • build_hip.bat: builds stable-diffusion.cpp with the HIP backend.
  • bench.ps1: a benchmark harness that times each sampling step.
  • docs/PITFALLS.md: more than 40 traps the author hit, from network issues to ROCm setup to measurement mistakes, plus deep dives on kernels and a memory-spill slowdown.
  • No model weights are included. The download script pulls the unsloth MiniMax-H3 GGUF files (the Q4_K pruned FL2VA model, a Q4_K_M Qwen3-VL text encoder, and the video and audio VAEs) from ModelScope. That's about 31 GB of disk.

Measured Presets

All of these were measured on the RX 6900 XT:

  • fast: 864×480, 22 frames (0.9 s of video), 8 steps, 67.2 s
  • draft: 512×288, 56 frames (2.3 s), 8 steps, 72.2 s
  • quality: 864×480, 22 frames, 20 steps, 138.5 s
  • longhd: 864×480, 56 frames (2.3 s), 8 steps, 183.0 s

You'll want 64 GB of system RAM, since the process sits at roughly 23–28 GB during a run.

The Fixes That Mattered

Windows silently spills big buffers

Once about 11.4 GB of weights are loaded, Windows quietly moves large working buffers (over ~600 MB) into shared system memory. Nothing in the logs says so, but some operations drop from 240 GB/s to 5–12 GB/s. The fix is to cap how much VRAM the weights can use:

--max-vram ROCm0=11

The cap has to be right on both sides: 11 GB gave 145 s, but 13 GB gave 400 s. The author's rule of thumb is to leave 4–5 GB free for working memory, so start at about VRAM minus 5 GB on other cards.

A smaller text encoder

Loading the 17 GB Qwen3-VL text encoder took 18 s on every run, whatever the prompt length. Switching to the 12.5 GB Q2_K_M quant cut that to 6.4 s, about 15% off the whole run, with no visible difference in side-by-side frames.

Two smaller wins

  • On the Vulkan backend, turning on flash attention made each sampling step 2.4× faster (94.1 s down to 38.9 s).
  • Raising one tile-size limit in ggml's RDNA2 kernel settings sped up the main matrix multiplications by 5.6–9.7%. The author thinks it's worth sending upstream to ggml.

The author also says the honest part out loud: the main math kernel already runs at 94% of what this card can do. Going much faster needs matrix cores, meaning RDNA3+ or NVIDIA tensor cores. More parameter tuning won't get you there.

Quick Start

.\scripts\download_models.ps1
.\scripts\extract_rocm.ps1
.\gen.ps1 -Prompt "A lighthouse keeper in a yellow raincoat climbs a spiral iron staircase at dusk, lantern swinging in her hand, rain streaking the round windows, waves booming below, the beam sweeping over a dark stormy sea" -Preset fast

The scripts want prompts of at least 25 words, and more detail helps. The download script saves to E:\minimaxh3\models by default, so edit $root first if your drive layout is different.

License Check

The scripts and docs are MIT, and stable-diffusion.cpp is MIT too. MiniMax H3 itself is not open source. It's under the MiniMax H3 Community License, which lists the US, EU, UK and South Korea as Excluded Territories that need a separate license. Read the model card before you download.

Original Source

Conclusion

This is one of my favorite kinds of drop. There's no new model here, just someone who refused to accept 62 minutes per clip and wrote down every wall they hit on the way to 3. If you've got a Radeon gathering dust, that --max-vram trick alone is worth the read. And if you'd rather skip the tuning entirely, H3 is ready to go on Gen.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=minimax-h3-amd-rx-6900-xt-windows-rocm-20x-faster

—Aurelia ♡