Weird name. But looks like it. Their blog even says “So easy to use, even a macaque could get the hang of it.” (source: https://huawei-csl.com/blog/macaq/ )
Why does some HN users seem to prefer GitHub page whenever available at expense of official site? I mean, Huawei is still sanctioned, anything they host third party outside China can disappear at any hour.
I prefer GitHub, because I can see the code and immediately know where to click. Product pages are always full of various marketing shots, videos, etc that just takes more time to parse through.
This is pretty much the reason that if I read about a new product, I generally just go to the Wikipedia page. I just want the no-bullshit version, no "Say hello to blah" video or marketing word salad.
The top 3 repos sorted by stars on Huawei's github page are all for quantization. I am not an LLM scientist so I can't tell what the difference is between them, or why they need three different pipelines for model quantization.
SINQ was created last year. It was for weights only and it was calibration-free quantization. Then KVarN is for KV-cache quantization to get longer context. MACaQ is multimodal, activation and weights calibrated quantization. 3 different projects
34 comments
[ 0.44 ms ] story [ 124 ms ] thread"Huawei releases MACaQ"
"Just like Jensen released his..."
Why does some HN users seem to prefer GitHub page whenever available at expense of official site? I mean, Huawei is still sanctioned, anything they host third party outside China can disappear at any hour.
Worth checking out beyond the GitHub
It is from their lab in Switzerland, after all.
Nice for Huawei to share it.