Home Creators Posts Import Register Favorites Logout

Content

The magic of the original FP16 14B model is now fully unleashed in Sugoi 14B Ultra! I have pushed the limits to retain almost every ounce of quality from the source model – and the result? Translation performance that’s ALMOST TWICE as good as the previous quantized Sugoi 14B version! (BLEU score of 21.38 vs 13.67)

But that’s not all – Sugoi 14B Ultra isn’t just accurate, it’s smart. With prompt-following skills rivaling the Qwen 2.5 base model, it’s ready to translate even text with lots of brackets (commonly seen in RPGM games) with unmatched precision.

Instructions:

https://blog.sugoitoolkit.com/sugoi-llm-14b-ultra/

Download links:

https://sugoi-file.sfo3.cdn.digitaloceanspaces.com/Sugoi-14B-Ultra-Q4_K_M.gguf

Comments

Gerald Gantos

Is there any chance we can get a less qauntized model q4 kinda hurts.

Amazing Flapples

you can quantize from the FP16 versions yourself with llama.cpp tool. I made a Q6 14B version like a month ago.

MingShiba

which quantization level do you want?

MingShiba

Can you post this on Sugoi Toolkit tech support channel along with some cmd screenshots, I'll have a look

J D

I'd like to try out a Q8 if possible please :D

Gerald Gantos

q8 is usually ideal balance for translating things.

coolguy11567

q6 also would be better anything is better than q4

Pilaxiv724

Are we able to get the non-quantized version? GGUF does not play nicely with certain hosts (particularly vLLM, which is blazing fast)

MingShiba

if there are more demands from users, I'll consider it because requests like you are right now less than 3 :)

Eyoldaith

+1 here, wanted to try making some legacy GGUF quants (Q4_0, Q6_0) of it to run on a tesla p40.