Llama 3

Llama 3 is Meta's flagship open-weights large language model family, released initially in 8B and 70B parameter sizes and later expanded with the Llama 3.1 405B model that closed much of the gap to closed-source frontier systems. The models are pretrained on roughly 15 trillion tokens, ship as both base and instruction-tuned variants, and support an extended 128K context window in the 3.1 generation. Llama 3 is the default starting point for anyone who wants a capable general-purpose chat or completion model they can actually run, fine-tune, and inspect themselves.

What sets Llama 3 apart is the license: weights are free to download under Meta's community license (with restrictions on services exceeding 700M MAU), which makes it the de facto open standard powering Groq, Together, Fireworks, Ollama, vLLM, and most fine-tuning pipelines on Hugging Face. There is no first-party hosted API priced per token from Meta itself — you either self-host or rent inference from one of the many providers that serve Llama-class models for cents per million tokens. That is the trade-off: zero vendor lock-in and full weight access, but you supply the GPUs or pick an inference partner.

For builders, Llama 3 is the workhorse behind countless RAG stacks, agent frameworks, and on-device assistants. Tooling around it (llama.cpp, Ollama, LM Studio, Unsloth, Axolotl) is the most mature in the open-source LLM ecosystem, and quantized GGUF builds of the 8B model run comfortably on a single consumer GPU or even a Mac.