I've switched to the iq3_xxs quant (published directly by stepfun-ai), and knowledge queries strangely seem to be even better at the lower quant. The model is also building software very well in Pi at this quantization. I haven't come across any loss of quality from the iq4_xs quant.

At iq3_xxs compression, the entire model (79.73Gb on disk) fits in memory on Strix Halo machines, with the full context length enabled. With that full context size, I haven't run into any crashes or weird results in Pi, which had been the case with iq4_xs quant.

Another benefit of the smaller quant is that the model runs faster at 27.77 tokens per second on Strix Halo. That's a very comfortable response speed for many tasks, and that's on a $2500 machine.

This model is getting to be one of my favorites for local inference. Qwen 3.6 35a3 with MTP is still my first go-to on basically every machine, including even those with inexpensive 3080 and 3080ti mobile GPUs, but for bigger machines, step-3.7-flash is enticing me more and more to actually use it on projects.