Shell script outruns torrent for 219GB LLaMA weights
It exists because the LLaMA torrent contained a forgotten direct download link that pulls 219GB of weights in 98 minutes instead of six hours.

What it does
llama-dl is a bash script that downloads all four LLaMA weight sets (7B through 65B), totaling 219GB. The author extracted a direct download URL buried in the original torrent metadata and wrapped it in a script that verifies file integrity against the torrent hashes. After Facebook killed the original leaked link within hours of the repo going public, the author mirrored the files to Cloudflare R2 and updated the script to point there.
The interesting bit
The script is a small act of digital archaeology: the torrent distributor had accidentally buried their own private Facebook download link in the metadata, which turned out to be roughly 3.6x faster than the swarm. The author verified the output by running webtorrent against the results and getting an immediate seed, confirming bit-for-bit identity.
Key highlights
- Downloads all weights in roughly 98 minutes at ~40MB/s, versus an estimated six hours via torrent
- Verifiable against the original torrent; files are identical down to the byte
- Includes a note about an improved sampler to fix Facebook’s poor default generation settings
- Mac users need a newer
bashfrom Homebrew because the script uses array syntax incompatible with Apple’s default shell - The author is transparent about the legal gray area of using a leaked private link, though the weights are now mirrored to R2
Caveats
- The original Facebook download link was shut off shortly after publication; the current script relies on an R2 mirror maintained by the author
- LLaMA is under a non-commercial bespoke license, so commercial use is legally risky regardless of how you obtained the files
- The author notes that early reports of LLaMA underperforming GPT-3 may be due to bad sampler settings, but also admits the 30B model failed to identify Captain Picard, suggesting it may not be OpenAI-grade
Verdict
Worth a look if you need the LLaMA weights quickly and don’t want to babysit a torrent. Skip it if you’re looking for model training code, an inference engine, or a legally unambiguous path to commercial use.
Frequently asked
- What is shawwn/llama-dl?
- It exists because the LLaMA torrent contained a forgotten direct download link that pulls 219GB of weights in 98 minutes instead of six hours.
- Is llama-dl open source?
- Yes — shawwn/llama-dl is open source, released under the GPL-3.0 license.
- What language is llama-dl written in?
- shawwn/llama-dl is primarily written in Shell.
- How popular is llama-dl?
- shawwn/llama-dl has 4.1k stars on GitHub.
- Where can I find llama-dl?
- shawwn/llama-dl is on GitHub at https://github.com/shawwn/llama-dl.