TL;DR: Llama is Meta's family of open-weight language models - you can download the actual model files, run them on your own hardware, and fine-tune them. That openness, more than any single release, is the family's identity: Llama kickstarted the open model ecosystem and remains one of its most common foundations.
What it is and how you use it
Llama is made by Meta, which took the opposite bet from OpenAI, Anthropic, and Google: publish the weights. Each Llama generation ships as downloadable model files in several parameter sizes, from small models that fit on a laptop to large ones that need serious GPU capacity. The license is a custom community license - free for almost everyone, with restrictions that keep it short of strict open source - which is why the careful term is open weights rather than open source.
The family's significance is the ecosystem it seeded. Because anyone can obtain the weights, a whole industry of derivatives grew on top: fine-tuned variants for specific domains, quantized versions compressed to run on consumer hardware, and research models that use Llama as a base. Tooling followed the same path - llama.cpp, the C++ engine that made local inference practical on ordinary machines, carries the family's name even though it now runs many model families.
Access is wherever you want it. Download the weights from Meta or Hugging Face and run them locally through Ollama or llama.cpp; serve them at scale on your own GPUs with an engine like vLLM; or skip self-hosting entirely and call a hosted endpoint from an inference provider or a major cloud. That flexibility is the practical case for the family: data never has to leave your infrastructure, costs are hardware costs rather than per-token fees, and fine-tuning on private data is fully in your control.
Where it typically fits: regulated industries and privacy-sensitive products that cannot send data to a third-party API, high-volume workloads where per-token pricing dominates the budget, and teams that want a model they can specialize. The trade-off is operational: you take on serving, scaling, and upgrades that a hosted provider would otherwise absorb, and the largest closed models usually stay ahead on raw capability.
Where it sits in the AI stack
Llama sits at the model layer, but unlike a closed family the serving layer is yours to choose - local runtime, self-hosted engine, or rented endpoint:
Key tools and implementations
-
Weight downloads
The model files themselves, distributed by Meta and mirrored with countless variants on Hugging Face.
-
Local runtimes
Ollama and llama.cpp run quantized Llama models on laptops and workstations with no cloud involved.
-
Serving engines
vLLM and similar engines turn your own GPUs into a high-throughput Llama API for production loads.
-
Hosted endpoints
Inference providers and major clouds serve Llama per-token for teams that want open weights without owning GPUs.
Related entries
- Open-weights models Models whose trained weights are published for anyone to download, run locally, and fine-tune.
- Local LLM Running a language model entirely on hardware you control instead of calling a hosted API.
- Mistral A French AI lab whose model family spans open-weight downloads and a commercial API, with a focus on efficiency.
- Fine-tuning Further training a pretrained model on your own examples to specialize its behavior for a task.