I'm not sure if this is specific to this quant, issues with the model, or a bug in either llama.cpp or openclaw, but I'm seeing laguna narrate the message, or analyze the metadata if sending from a messaging platform. For me, it doesn't seem to be working very well with either openclaw or hermes.
So the i-mini is fast but admittedly a little lobotomized. I had another person mention that the mini was a little flaky. I can experiment with it a little more. If you're using that, you're probably going to need to drop the temperature a little bit (which narrows the distribution), and tighten the top_p and and min_p which chops off the tails. I'd try <0.7 for the temperature, 0.05 to 0.1 for the min_p and/or drop the top_p to 0.9 or even a little less. This chops off the tails of the distribution. For code you can go very low on the temperature. A lot of the SWE bench folks go down into the <0.15 range to get the high benchmarks they are reporting. I haven't seen whether that's what Poolside is doing but it's worth a shot.
Check the other poster's comments for his settings, especially the dry base and multiplier. I honestly have only extensively run the i-quality versions except for my benchmarks because I have the headroom.
FYI, I'm planning to do an intermediate quant between mini (IQ2_S mid experts) and compact (Q3_K) with what I'm calling subcompact and using Q3_XXS mid experts to get a little more diversity at the lower end and be ~1/2 way between the two in terms of size. I took the "mini" recipe and bumped up the lowes tier to Q3_XXS, purely based on size in GB to fall in the middle.
