Introduction

MiniMax H3 is the open audio-video model everyone wants to run locally, and the out-of-memory error is the wall most of us hit. MiniMax H3 LowVRAM is a new GPL-3.0 tool by developer jiangzechunwnl-source. Version 0.1.0 went up on GitHub today, Nov 1. It's a small offline launcher that patches native ComfyUI's H3 implementation to use far less activation memory while still computing full attention.

It comes with a double-click Windows setup ZIP and a manual launch path for Windows or Linux. It adds no extra Python packages, no model downloads and no cloud service. Your normal ComfyUI install is left untouched.

What Shipped

How it saves memory

  • It projects QKV and computes the MLP in 1024-token chunks, writing results back into the rows it has already consumed.
  • It builds the complete K/V first, then recomputes Q chunk by chunk against all of it. That keeps the math full-attention, not a local-window approximation, and full Q is never held in memory.
  • Each layer's converted or LoRA-merged weights are reused across its chunks.
  • Video and audio embedding storage is freed once it has been copied into the packed sequence.

The headline number

On Linux, an H3 INT8 + Turbo 8-step request at 768×1344, 10 seconds, with audio had previously run out of memory on an 8 GB RTX 5060 Laptop. With the patch, sampling went through, and the logged PyTorch allocator peak was 4,655 MiB. The author also reports successful runs at larger resolutions afterward.

Read that number carefully, though. It's an allocator peak, not total GPU memory, and the recorded first step took about 500 seconds. This trades time for memory. The author explicitly says it isn't a speed-up, and a 4 GB card is not guaranteed to work.

Safety Rails I Like

  • Fingerprint locked: it supports exactly one verified native H3 source baseline (ComfyUI commit dd79c643). It checks that fingerprint before launch and refuses to run on anything else.
  • Isolated launch: it starts ComfyUI with --lowvram, --cache-none, a 1 GiB VRAM reservation, and all custom nodes and API nodes disabled. Use a native H3 workflow.
  • Offline by design: external connections and child processes are blocked by the launcher.
  • Easy undo: close it and use your normal launcher again. Setup only adds a folder and a start script.
  • Don't stack it with KJNodes H3 optimizations, LongMedia, Sage or Sol patches, or other weight-cast swaps.

Try It

On Windows, download the release ZIP, extract it, run Install-Windows.cmd, and pick the main.py inside your ComfyUI folder. Then close your normal ComfyUI and start Start-H3-LowVRAM.cmd. On Linux or for a manual launch, use the Python that already runs your ComfyUI:

python run.py --comfy-root /path/to/ComfyUI --check
python run.py --comfy-root /path/to/ComfyUI

It serves on 127.0.0.1:8188 (add --port 8189 if that's taken). Start with a short, small request and scale up gradually.

Honest Caveats

  • The checks are small-layer numerical comparisons in FP32 and BF16 at 1024, 1025 and 2051 tokens, and all nine combinations passed. That isn't a full-checkpoint or final-video comparison, and bit-identical output isn't promised.
  • Full standalone video generation on other machines, GPU architectures and quantization backends is still unverified.
  • Weights, text encoding, VAE decoding and system RAM still need their own headroom, so another stage might become your new bottleneck.

Original Source

Conclusion

Not every breakthrough is a new model. Sometimes it's someone patiently getting the model you already have to fit. This one is experimental, fenced in and careful about its claims, and I love that. If your 8 GB card keeps hitting the out-of-memory wall on H3, it's worth an evening, a native workflow and a cup of tea while that first step runs. And if you'd rather skip the waiting entirely, H3 is ready to go on Gen.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=comfyui-h3-lowvram-minimax-h3-8gb-chunked-attention

—Aurelia ♡