LLaMA

lastmod 2026-02-23created 2023-03-0233 edits

A 65b-parameter LLM released by Meta

Cateogries

• Llama 2 - Meta releases its next-generation language model Llama 2
• Llama 3 - Meta releases its next-generation language model Llama 3
• Llama 3.1
• Llama 3.2 - Revolutionizing edge AI and vision with open, customizable models (Meta released the first large multimodal model (LMM) in its 'Llama' series that understands both images and text)
• Llama 3.3 - with 70B, a GPT-4-class model can now run on a laptop
• Llama 4 - the most Korean-friendly open-source model.
• nano-llama31 - a nanoGPT-style version of Llama 3.1
• ntransformer - an NVMe-to-GPU inference engine that runs Llama 3.1 70B on a single RTX 3090

Features

• Released in 4 sizes: 7B, 13B, 33B, and 65B
• Much smaller in scale, but an efficient model comparable to larger models thanks to more extensive data training and fine-tuning
• 33B/65B were trained on 1.4 trillion tokens (7B on 1 trillion)
• "The 13B model outperforms the 175B GPT-3, and 65B is competitive with the much larger Chinchilla70B and PaLM-540B"
• Can only be used for non-commercial purposes such as AI research (requires applying and getting approval)

Downloads

• Meta's LLaMA model is being distributed via torrent | GeekNews
• Github - shawwn/llama-dl: High-speed download of LLaMA, Facebook's 65B parameter GPT model

Linux:

curl -o- https://raw.githubusercontent.com/shawwn/llama-dl/56f50b96072f42fb2520b1ad5a1d6ef30351f23c/llama.sh | bash

For reference, when I downloaded it on answerver, it took about 157 minutes and used 220G of space.

See also

• Meta (Facebook)
• GPT-3
• ChatLLaMA - an open-source implementation of ChatGPT based on LLaMa
• llama.cpp - inference of Facebook's LLaMA model in pure C/C++
• Alpaca - a powerful open-source Instruction-Following model at the GPT-3.5 level
• Dalai - the simplest way to run LLaMA on your computer
• Code Alpaca - LLaMA trained on code generation instructions
• Simple LLaMA Finetuner - a simple, intuitive UI tool that lets even beginners easily fine-tune LLaMA-7B
• Lit-LLaMA - an open-source LLaMA implementation based on nanoGPT
• GPT4All - a chatbot based on LLaMA, trained on an 800k instruction set
• xturing - build and control your own LLM
• Koala - a dialogue model for academic research (a LLaMA-based chatbot)
• RedPajama - an open-source project that recreates the LLaMA dataset
• llama-tokenizer-js - a LLaMA tokenizer that runs in the browser
• Lag-llama - the first open-source foundation model for time series forecasting
• Llama.vim – a code/text autocompletion plugin for VIM using LLMs
• LlamaFirewall - a framework for detecting/blocking AI security risks

Favorite site

• Introducing LLaMA: A foundational, 65-billion-parameter language model
• LLMs are having their Stable Diffusion moment | GeekNews
  • [Original] Large language models are having their Stable Diffusion moment
• A brief history of LLaMA models | GeekNews
  • [Original] A brief history of LLaMA models - AGI Sphere
  • LLaMA - (7B, 13B, 33B, 65B), CommonCrawl/C4/GitHub/Wikipedia/Gutenberg & Book3/ArXiv/StackExchange
  • Alpaca - 52k GPT-3 instructions
  • Vicuna - 70k ChatGPT conversations
  • Koala - 117k cleaned ChatGPT conversations
  • GPT4-x-Alpaca - 20k GPT4 instructions
  • WizardLM - 70k instructions synthesized with ChatGPT/GPT-3
  • OpenAssistant - 600k human interactions (OpenAssistant Conversations)