01

Size and quality are normally a trade

Quantizing a model smaller almost always costs you something in quality. The usual question is just how much you're willing to pay. ROCmFPx, the 4-bit floating-point format we built for Strix Halo's unified memory (see the adaptive speculation writeup for the rest of that fork), was built to bend that trade rather than just ride it.

We compared our ROCmFP4-FAST-imatrix build of Qwen3.8-Flash-Next against the leading community IQ4_XS requant of the same model, from quantizer AesSedai.

02

Smaller, not worse

BuildSizeΔ perplexity vs. reference
AesSedai IQ4_XS 117.13 GiB +3.12%
Agention ROCmFP4-FAST-imatrix 87.06 GiB +2.48%

30.07 GiB smaller, with a smaller perplexity penalty too: +2.48% versus +3.12% relative to the unquantized reference. A quant this much smaller usually costs quality; this one doesn't.

03

Get it

The build is sized for Strix Halo's 96GB unified memory and runs entirely on GPU. Full benchmark methodology and download on HuggingFace, or see it alongside our other builds on the models page.

View on HuggingFace ↗