Llama 3

Llama 3 — explained in more detail

Llama 3 is the model generation Meta released on 18 April 2024 and a central example of open-weight language models. Meta made the weights available for download — usable for research and, under the Llama 3 community license, for commercial applications too. At launch, Meta positioned Llama 3 as the most capable openly available LLM at the time.

The first release included two sizes: 8 billion and 70 billion parameters, each as a pretrained base and an instruction-fine-tuned (Instruct) variant. Both use Grouped-Query Attention (GQA) for more efficient inference. The context window of the first Llama 3 models was 8,000 tokens. Meta already announced larger models (over 400 billion parameters) and future capabilities such as multimodality, more languages and longer context — these followed in later point versions (Llama 3.1, 3.2, 3.3).

The models are available from Meta itself as well as via platforms such as Hugging Face and major cloud providers.

Example / practical relevance

Llama 3 became the foundation of many open AI applications: the 8B model runs on comparatively modest hardware and suits local assistants, chatbots or fine-tuning on custom data; the 70B model delivers higher quality for more demanding tasks.

Because the weights are open, companies can self-host and fine-tune Llama 3 — relevant for data protection, cost control and independence from an API provider. Note the community license: it is permissive but includes conditions (for example above very high user counts).

Distinction from similar terms

Llama is Meta’s model family and stands alongside other open families such as Mistral, Qwen (Alibaba) or DeepSeek. Llama 3 is an older generation; later versions (Llama 3.1 with long context and a 405B model, Llama 3.2 with vision, Llama 4) built on it. The Llama community license is not identical to classic open-source licenses such as Apache 2.0 or MIT.

Discover more

Blog

GPT-6 Astra Found Questions My First Security Review Missed

I used GPT-6 Astra and Fable 5.1 as independent reviewers for RLS, API and tenant-isolation checks. The useful part was the structured cross-review.

Glossary

Code Llama

Code Llama is Metas coding-specialized Llama 2 variant from August 2023 — open-weights, in several sizes and with base, Python and Instruct flavors.

Encyclopedia

Comparing AI models — who builds what and how to choose

The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.

Topic overview

AI ModelsLlama