Qwen 3.5 VS Qwen 3

Since I've been spamming locallama with it anyways, here's my config that prevents loops. One thing wrong and it loops. 3 days of tests with all kind of different quants from different vendors, different settings. Bigger variants might be less senstive. \- you should use q8\_0 quants \- use bf16 for kv cache, qwen3.5 is very sensitive to it because of it's architecture \- adapt -t for number of physical cores, \- enable flash-attn for modern cards (me -> old rtx2060), \- increate -b and -ub for modern cards \- config might not be best for non-thinking mode ./llama-server \\ \-hf bartowski/Qwen\_Qwen3.5-2B-GGUF:Q8\_0 \\ \-c 92000 \\ \-b 64 \\ \-ub 64 \\ \-ngl 999 \\ \--port 8129 \\ \--host [0.0.0.0](http://0.0.0.0) \\ \--flash-attn off \\ \--cache-type-k bf16 \\ \--cache-type-v bf16 \\ \--no-mmap \\ \-t 6 \\ \--temp 1.0 \\ \--top-p 0.95 \\ \--top-k 40 \\ \--min-p 0.02 \\ \--presence-penalty 1.1 \\ \--repeat-penalty 1.05 \\ \--repeat-last-n 512 \\ \--chat-template-kwargs '{"enable\_thinking": true}'