I have used qwen 3.5 for several months, the albond and prismaquant variants.
I use it to develop Java software in OpenCode.
I tried the ds4 version from entrpi for a few days, for planning a larger task (new Bounded Context).
It is a bit too early to say, but it seems that the 0xSero variant is the currently best one for my use case.
I am however closely following gridbook. Soon it should also support DeepSeek. Exciting new Enhancements to PrismaQuant -- GridBook - #73 by tenari .
It seems to me that the focus is too much on tokens/second. I’m only interested in two things: total time needed for the task and the quality of the output. If the model does “35 tok/s”, but keeps looping/overthinking, the task takes longer.
Have a look at C1 1058pp/s, 52 tg/s on 1x dgx spark on Deepseek V4 flash 0731 full 256 experts - #15 by siertum