I've been running a lot of experiments, from repeat yourself to extending ngram mod to store the ngrams across conversations and reboots, at least under the conversation that I have been replaying again and again and again... the q4s head behaved JUST as the q6 and f16 regarding acceptance rate, and it's faster while giving you back ram.
This is using the https://huggingface.co/mradermacher/Qwen-3.8-27B-Uncensored-... quants.
I've been running a lot of experiments, from repeat yourself to extending ngram mod to store the ngrams across conversations and reboots, at least under the conversation that I have been replaying again and again and again... the q4s head behaved JUST as the q6 and f16 regarding acceptance rate, and it's faster while giving you back ram. This is using the https://huggingface.co/mradermacher/Qwen-3.8-27B-Uncensored-... quants.
[dead]