Llama
Llama is Meta's open-source large language model family, currently the benchmark for open-source LLMs. The latest Llama 3.x series approaches closed-source flagships on multiple benchmarks.
Version history
- Llama 1 (2023.2): first open release, research-only license.
- Llama 2 (2023.7): commercial use allowed, 7B/13B/70B tiers.
- Llama 3 (2024.4): 8B/70B, ~15T training tokens.
- Llama 3.1 (2024.7): 405B flagship, the largest open model.
- Llama 3.2 (2024.9): 1B/3B on-device + multimodal variant.
Why Llama matters
- Best ecosystem support: vLLM, TGI, llama.cpp, ollama, Transformers all first-class.
- Quantization community benchmark: GGUF, GPTQ, AWQ all prioritize Llama support.
- Most derivative models: Mistral, Qwen, DeepSeek all trained on Llama architecture.
- Largest fine-tune community: tens of thousands of fine-tunes on HuggingFace.
Key improvements
- Llama 2 → 3: tokenizer significantly better (Chinese tokenization 4x more efficient), GQA.
- Llama 3 → 3.1: context from 8k to 128k, 405B approaching GPT-4.
Commercial terms
Llama Community License allows free commercial use under 700M MAU; above that needs separate licensing.