Meta’s official Llama utility belt: download, quantize, and run

What it does This is Meta’s companion package for its Llama model family. It provides a llama-model CLI to list, download, verify, and remove model weights from Meta’s own servers or Hugging Face, plus reference inference scripts for Llama 3 and Llama 4. Think of it as the official plumbing between the model weights and your GPUs.

The interesting bit The repo doubles as a memory-footprint cheat sheet. The README spells out exactly how many GPUs you need for Llama 4 at full precision (four), with FP8 mixed quantization (two 80 GB cards), or Int4 mixed (one 80 GB card). That level of frank hardware admission is rare in official model repos.

Key highlights

• CLI supports downloading from Meta (via signed URLs after license acceptance) or Hugging Face, with integrity verification built in.
• Reference chat and completion scripts for Llama 3 and Llama 4 live under models/<version>/scripts/.
• Quantization modes (fp8_mixed, int4_mixed) are baked into the inference examples to shrink GPU requirements.
• Catalogs the full family tree: Llama 2 through Llama 4, including context lengths up to 10 million tokens for Llama 4 Scout.
• Distributed as a PyPI package and also integrates with Hugging Face transformers.

Caveats

• Downloading weights requires accepting a license on Meta’s website and waiting for an email with a signed URL that expires in 24 hours.
• Llama 4 inference at full precision demands serious hardware; the README notes at least 4 GPUs.

Verdict Grab this if you need the official, no-surprises path to downloading Llama weights or want Meta’s own baseline inference scripts. Skip it if you already manage models through Hugging Face or llama.cpp and don’t need the reference runner.