Torrents are perfect for model distribution

pogue

I feel like if Bittorrent shipped in the browser, either embedded or as an extension and just ran on it's own accord, managing the download and throttling upload & download speed based on my current usage and needs without a lot of tweaking and configuration setup, that would be fine. I know most browsers tried this and gave up on it years ago. But, I'd rather not manage a separate program, having to fiddle with settings and open ports and have it always running in the background using a lot of ram/cpu, it's just a hassle. But, are big CDNs really struggling to manage people downloading 500+gb models? Is AWS/Cloudflare/Akamai/etc begging for p2p hosting again?

Havoc

Torrent always made more sense for distributing large static files like models especially ones where specific ones suddenly become popular and need scaling. Huggingface is incentivised to not do that though to keep control. Wish they'd at least gone for a hybrid approach

alicehuang1

DHT/torrents and AI have been on a collision course for the last 20 years. Sometimes the correct answer is the simplest one. Except the stakes are much higher now with the power struggle over frontier models.

unscaled

> Movies are different. Search for a movie on a torrent site and you'll see 4K remux, 1080p, 720p [...]. But no one downloads models like this: anyone who wants DeepSeek-V4.1-Flash-1017 wants the same 48 shards. But this is exactly the situation for AI models as well. If you look for DeepSeek-V4.1-Flash-1017 or Qwen3.8-27B on HuggingFace, you'll see many different releases by various people with different quantizations like FP8, Q4_K_M, INT4, abliterated/uncensored models and various versions that are fined tuned or modded in whichever different way. It feels to me exactly like the various different video encodings. Yes, there is usually a single, official release from the lab that produced the model (or sometimes a handful of different variants), but very frequently this isn't something you can even run on consumer hardware.

flyby25

Very few users will end up downloading the base model straight from the lab, chief reason being they can't actually use it for inference on consumer hardware. We will end of with the same scenario described for films where lots of competing providers for different quantisations and fine-tunes will dilute the swarm. I can only assume this was written from the perspective of a developer/researcher?

blosko

this make a lot of sense, i wonder why hugging face didn't offer this off the bat. makes you wonder