hunyuan does okay on mesugaki text
grab iq4xs from https://huggingface.co/qwp4w3hyb/Hunyuan-A13B-Instruct-hf-WIP-GGUF
git clone https://github.com/ngxson/llama.cpp
cd llama.cpp
git fetch origin pull/26/head
git checkout -b pr-26 FETCH_HEAD
./llama-server --ctx-size 4096 -b 1024 --jinja --no-warmup --cache-type-k q8_0 --cache-type-v q8_0 --flash-attn --temp 0.6 --presence-penalty 0.7 --min-p 0.1 --model ~/TND/models/hunyuan-a13b-instruct-hf-WIP-IQ4_XS.gguf -ot exps=CPU -ngl 99 --no-mmap

prompt eval time = 1893.24 ms / 25 tokens ( 75.73 ms per token, 13.20 tokens per second)
eval time = 132688.70 ms / 874 tokens ( 151.82 ms per token, 6.59 tokens per second)
total time = 134581.93 ms / 899 tokens
vram usage: ./llama-server 3190MiB
>captcha: 4080D