<h1 id="minimax-h3-w4a8-quantizer">MiniMax H3 w4a8 Quantizer</h1>(Directions are under the cut!!)Observation: Kijai's <a target="_blank" rel="ugc" href="https://huggingface.co/Kijai/MiniMax-H3-experimental/tree/main">w4a8 version of H3</a> is excellent; very close to identical to the int8_convrot at half the memory cost.Problem: People aren't making w4a8 versions of their H3 finetunes!Solution: You're looking at it!This is a Python script that converts a full-precision BF16 file for a Minimax H3-based model, and spits out a w4a8 version. It streams the weights from disk layer by layer, instead of loading the entire model at once, so you DON'T have to have 70GB of memory available to use it! (If you can use Kijai's w4a8 of the base model, you should be good to go.)The script itself was written by GPT 5.6 Terra, but I tested it myself both making a pruned and unpruned version and it works great. The first video here is with my w4a8 quant of PinkCherry; the second is the official int8_convrot.<h2 id="directions"> Directions</h2><ul><li>Download this archive and extract it somewhere.</li><li>Download and install <a target="_blank" rel="ugc" href="https://docs.astral.sh/uv/">uv</a>, or if you don't want to do that, see "Python venv" below. (Use uv though.)</li><li>Download a full-precision BF16 MiniMax H3 model. To replicate the videos above, use <a target="_blank" rel="ugc" href="https://civarchive.com/models/2838593/pinkcherry-mm-h3-fl2va">PinkCherry beta 0.6.</a> Put it next to the script. (I haven't tested it with things that alter the model architecture in any arcane way like 10Eros. You can try though. Report back!)</li><li>Open a terminal in the same directory as the script and the BF16 weights and run this, replacing &lt;MODEL&gt; with the name of the BF16 .safetensors file and &lt;QUANT&gt; with... whatever you'll remember.</li></ul><pre><code>uv run --script convert_minimax_h3_w4a8_streaming.py &lt;MODEL&gt; &lt;QUANT&gt;</code></pre>By default, this script keeps <code>adaln</code> intact, but if you want to make a pruned checkpoint, you can pass <code>--prune-adaln</code>, like so:<pre><code>uv run --script convert_minimax_h3_w4a8_streaming.py --prune-adaln &lt;MODEL&gt; &lt;QUANT&gt;</code></pre>To create the file used in the example videos, I did this (the backslashes are just so you can use newlines for readability):<pre><code>uv run --script convert_minimax_h3_w4a8_streaming.py --prune-adaln \
 PinkCherry_fl2va_MiniMax_H3_bf16_beta-0.6.safetensors \
 PinkCherry_fl2va_MiniMax_H3_w4a8_pruned_beta-0.6.safetensors</code></pre>Once finished, put the resulting file in your <code>diffusion_models</code> folder, and give it a test! If it's to your liking, you can delete the BF16 weights to reclaim your disk space.(Note that I haven't tested this with things that alter the model architecture in any arcane way like 10Eros. You can try though. Report back!)<h3 id="python-venv">Python venv</h3>If you don't want to/can't install uv for some reason, the script can also be used with Python &gt;= 3.10, with these dependencies. This is the "I know what I'm doing" path, please don't ask for help with this, I use uv to not have to deal with this stuff!<ul><li>numpy&gt;=1.26,&lt;3</li><li>torch==2.7.1 (if you need CUDA then make sure to pass the right <a target="_blank" rel="ugc" href="https://download.pytorch.org/whl/cu128">https://download.pytorch.org/whl/&lt;cuXYZ&gt;</a> index URL as well)</li><li>triton==3.3.1</li></ul>Video made from a Krea 2 image I made while shitposting with a friend. There's your explanation for it.

Minimax H3 W4A8 Quantizer

Rewritten by GPT-5.6 Sol. It behaves exactly the same as the original, except it isn't a complete travesty to read, and the error behavior was improved to fail explicitly instead of quietly producing a broken output. It doesn't really matter which one you use, the model it produces is byte-identical.

Files with SHA256: 98f4d7e57c8d86e67ae6494783ab534a1600f3b23b834b968db75cbae3bf4cb5