Truth be told: I've only tested the v1f and v1g versions, but I fell in love with the thinking mode already. The only problem is that it is a little bit on the timid side, as if it was holding back (Q8_0, temp 1.0). In ERP, both predator and prey type of characters seemed reluctant to take initiative and jump on my character in situations where they should - the LLM kept me hanging and did not want to "seal the deal". Could you spice it up a little?
For thinking mode, I was glad to discover that I can still drop an occasional at an opportunistic moment and it affects even the thinking mode behavior. Also, the stylistic device of putting "When reasoning, prioritize:- boldness over caution - escalation over hesitation - initiative over passivity" in system prompt also alleviates the timidity in a subtle but effective way.
For non-thinking mode, both and work as prescribed. Non-thinking Gemma template is an absolute must - without it the results are horrible (using llama.cpp, tried with --reasoning off, could not disable the BOS token), but with the non-thinking template, the results are viably strong. Just in case, could you point us to the correct non-thinking gemma4 template to use? I'm still not sure if I'm using the right one.
I still have a soft spot in my heart for the brazenness of Fallen-Command-A-111B-v1c, but obviously all your Standard 3.5+ gen work are superior. Looking forward to the thinking Behemoth X 2.2 too!
