GGUF metadata clarification on some Qwen3.5 models
Hey Bartowski! First of all, highly appreciate the quants you've been putting out for quite some time, glad there have been many folks to fill the shoes TheBloke left behind 😄
I've been downloading the various small Qwen3.5 models from your repos to try to find the best sizes and quant mixtures for my random devices (anything capable of running llama.cpp lol), checking the ggufs' metadata tab to figure out which mixtures of quant types seem to perform best on which computers, and noticed that on the 35B, 9B, and 4B models the "base" model name and url seem to point to Qwen3.5 Base instead of just Qwen3.5 like on the Unsloth quants.
I didn't check all the quants and only noticed it on these smaller models (2B & 27B+ didn't have the Base model name that I could find). Also not sure if this actually means they were converted from the wrong repo or if the metadata was just saved incorrectly, but wanted to bring it to your attention in case it mattered.
Thanks!
Great info thank you sir!
For what it's worth, I've been using the Q8_0 of this model (35B) since yesterday for various conversations and web searches in openwebui and haven't had any moments that felt "off" about the intelligence, looping reasoning, etc. But I've been focused a lot more on testing various combinations of llama-server args to see how fast I can get them running vs the output quality.
I don't modify those manually, I believe it's just been automatically loaded into the model metadata, perhaps by parsing the README.md file from the original model which contains
base_model:
- Qwen/Qwen3.5-35B-A3B-Base
You can trust what you see in the side bar on huggingface :) that metadata I DO set myself and it will be accurate!
This is indeed not the base model and is from the instruct tune, but good looking out !
Thanks makes sense! Yeah I've been comparing your 4B with Unsloth's and hadn't noticed any real differences there either 👍