Let's optimize nvidia/GLM-5.3-Flash-NVFP4 for 2× DGX Spark

turboderp did make a rather amazing K3 quant of DSV4F. But I’m a bit wary. The current EXL3 version of this by brandonmusic (used by entrpi and miaai lab atm) is just really intelligent. I wonder how it tests comparatively.

I have seen looping once with an NVFP4 quant for what it’s worth. Have not seen it with EXL3.

This fellow here did apparently try to optimize cache size on EXL3:

I haven’t tried it yet. Let me know if you do.