Hacker News Reader

Comment by chorizo | Hacker News Reader

If HF didn’t exist, model releases would likely have been on torrents like Linux iso’s back in the day.

I'd honestly prefer that. I'm usually on 5g if not 4g and pulling models is an absolute nightmare when I basically have to use HF and even with a token getting throttled. Although I have a pretty unique use case. Pulling wads of 5gb tensor files over cellular is painful.

I am on the same boat, I live in a small Himalayan village. I have 2x5G based devices and one Wireless bridge (Ubiquiti LiteBeam M5) for a local broadband. Generally I get 50 Mbps, sometimes up to 100 Mbps

i'm traveling in an RV and can always (80% of the time) get up to 3MBps, max to 300. But with cellular it's mega spotty so pulling 3-10 5GB files then losing your connection for a few seconds/minutes completely kills that files download. I've tried about 10 different keep-alive-ing type hf downloader clients and with my 64gb ram I'm usually downloading 25-32GB models and I don't even try unless I'm in sitting in a city next to a library anymore.

It's less of a pain now that I'm not in my LLM newbie phase of trying to test out 30 different models and stick with Qwen, but I couldn't even try to download qwen 3.8 for a week or two until I got back to the city.

Starlink is the obvious option and I have it on my roof but I do everything I can to not pay that $170 a month vs my $170/quarter 5g service.

Reply on news.ycombinator.com