Qwen3.8-27B: quants compared
The same model in 19 variants, largest file first. The smaller the file, the fewer bits are left per weight. The full-precision variant ran in the cloud, all others on the Mac Studio. Click a picture to open the giraffe.
UD-Q4_K_MUnsloth Dynamic 3.1On the main page
16.6 GB≈ 4.9 bits per weight
UD-IQ3_SUnsloth Dynamic 3.0
12.0 GB≈ 3.6 bits per weight
UD-IQ3_XXSUnsloth Dynamic 3.0
10.9 GB≈ 3.2 bits per weight
piNo working giraffeNo page: in two attempts the model used up the 32,768-token output limit, mostly on thinking, and wrote no index.html.
UD-Q2_K_XLUnsloth Dynamic 3.1
9.9 GB≈ 2.9 bits per weight
UD-IQ2_SUnsloth Dynamic 3.1
8.4 GB≈ 2.5 bits per weight
UD-IQ2_XXSUnsloth Dynamic 3.1
7.2 GB≈ 2.1 bits per weight
OpenCodeNo working giraffeEmpty page: a call to a function that does not exist in Three.js stops the script before the scene appears.
UD-IQ1_MUnsloth Dynamic 3.0
6.7 GB≈ 2.0 bits per weight
piNo working giraffeNo page: the model wrote giraffe.js but reported done in two attempts without an index.html.
OpenCodeNo working giraffeWhite page: loads Three.js as a local ES module, which the browser blocks when the file is opened directly; the page also never attaches the canvas.
UD-IQ1_SUnsloth Dynamic 3.0
6.2 GB≈ 1.8 bits per weight
piNo working giraffeNo page: in two attempts the model thought at length and then stopped without writing a file.
OpenCodeNo working giraffeNo page: the model ended both attempts after one reply without writing a file.
































