The history of Llama, from LLaMA to Llama 4

Llama in one table

Meta released its first Llama model in February 2023 and its latest Llama generation, Llama 4, in April 2025. In between came five more releases, each pushing size, context length or openness a step further. The table below lists every major release; the sections after it explain what changed and why it mattered. Dates and sizes come from Meta's announcements and model cards.

Four of these models are in our chat today: Llama 4 Maverick, Llama 4 Scout, Llama 3.3 70B and Llama 3.1 8B. The older releases and the newer Muse models described below are not in our chat.

2023: from research release to open weights

LLaMA (February 2023)

The first model, written "LLaMA" for Large Language Model Meta AI, came in four sizes from 7 to 65 billion parameters. Meta presented it as a foundation model for researchers: smaller than the largest systems of the time, so that people without huge computing budgets could study it. Access was under a non-commercial license and granted case by case to academics, government and civil society groups, and industry labs. It was the start of a line that would soon open up much further.

Llama 2 (July 2023)

Llama 2 was the turning point. Meta released it in 7B, 13B and 70B sizes, together with chat-tuned versions, and made it free for both research and commercial use. Microsoft was named as a preferred partner. The context window was 4K tokens. With Llama 2, an open-weight model became a practical option for companies, not only for researchers.

2024: bigger, longer, multilingual

Llama 3 (April 2024)

Llama 3 arrived in 8B and 70B sizes, each as a pre-trained base model and an instruction-tuned chat version, with an 8K context window. It set the pattern for the point releases that followed during the year.

Llama 3.1 (July 2024)

Llama 3.1 extended the context window to 128K tokens, added official support for eight languages and introduced a 405B model, by far the largest openly available Llama. The license also explicitly allowed using Llama's outputs to improve other models, for example to generate training data. Llama 3.1 8B is still one of the most popular small models for running on a laptop.

Llama 3.2 (September 2024)

Llama 3.2 went small and visual at the same time. The 1B and 3B text models were built for phones and other devices, while the 11B and 90B Vision models could read images for the first time in the Llama line. The license for the vision models came with a restriction for people and companies based in the European Union.

Llama 3.3 (December 2024)

Llama 3.3 was a single 70B text model. It kept the 128K context and eight languages of Llama 3.1 and gave developers an updated mid-sized model. It remains a common choice for text tasks today.

2025: Llama 4 and the unreleased Behemoth

Llama 4 Scout and Maverick (April 2025)

On April 5, 2025, Meta released Llama 4 Scout and Llama 4 Maverick, its first Llama models built as a mixture of experts. Both have 17 billion active parameters per token; Scout has 109 billion in total across 16 experts, Maverick 400 billion across 128 experts. Both read images natively, both officially support 12 languages, and Meta listed context windows of 10 million tokens for Scout and 1 million for Maverick. Their training data ends in August 2024.

Around the same time Meta published Llama Guard 4, a 12B safety classifier derived from Scout, for filtering prompts and answers in apps.

Llama 4 Behemoth (never released)

At the Llama 4 launch, Meta previewed a much larger model called Llama 4 Behemoth and said it was still training. In May 2025, Reuters reported that its release had been delayed. Meta never published it. Claims of a "Llama 5" that circulate online are not backed by any Meta release: as of October 2026, Llama 4 is the newest Llama generation.

A new lab

In 2025 Meta reorganized its AI work under a new group, Meta Superintelligence Labs. That group went on to build Meta's next models under a new name.

2026: Meta moves on to Muse

Meta's newest models are not called Llama. None of them are available in our chat; this section is here so you know what exists.

• Muse Spark (April 8, 2026) was the first model from Meta Superintelligence Labs. It powers the Meta AI app and meta.ai. Meta did not release its weights and offered it to selected partners through an API preview.
• Muse Glimmer (August 10, 2026) is a 30-billion-parameter open-weight model aimed at agent tasks, released under the Apache 2.0 license.
• Muse Wonder and Muse Wonder Full (October 12, 2026) are Meta's most capable models so far. Wonder is offered through Meta's Model API; Wonder Full is limited to paid subscriptions. Neither has downloadable weights.
• Muse Spark Open (October 15, 2026) is an open-weight version of Spark under a new Muse Model License, which Meta describes as having limited commercial terms.

Two changes followed for Llama users. First, Meta's assistant moved from Llama to Muse. Second, Meta closed its hosted Llama API preview on July 6, 2026. Meta's developer help center now says Llama models are not served through the Meta Model API; Llama is something you download and run yourself or use through other providers. Our Llama API guide explains the options that still work.

What stayed the same

Through all these changes, a few things have been constant. Meta publishes Llama weights for anyone to download after accepting a license. Each version has its own community license with similar core terms: commercial use is allowed, products must credit "Built with Llama", an acceptable use policy applies, and companies with more than 700 million monthly active users need a separate license from Meta. Our license page explains these terms in plain language.

The models themselves also stay available. Llama 4, Llama 3.3 and Llama 3.1 are still offered by many cloud providers and run on home computers with tools such as Ollama; see run Llama locally. And because the weights are open, anyone can keep building fine-tuned versions on top of them.

What this history means for you

If you just want to chat, the history mostly explains the model menu. Maverick and Scout are the newest Llama models and the only ones in our chat that read images. Llama 3.3 70B is the strongest text-only option, and Llama 3.1 8B is the quickest. For help picking one, read how to choose the right Llama model, or start with the default:

Chat with Llama 4 Maverick

FAQ

What is the newest Llama model?

Llama 4 Scout and Maverick, released on April 5, 2025. Meta's later models are called Muse, not Llama.

Is there a Llama 5?

No. As of October 2026, Meta has not released a Llama 5. Its newer models carry the Muse name.

Was Llama 4 Behemoth ever released?

No. Meta previewed it in April 2025 as still in training, its release was later reported as delayed, and it was never published.

Can I still download older Llama models?

Yes. Meta keeps Llama 2, 3, 3.1, 3.2, 3.3 and 4 on Hugging Face under the meta-llama organization. You accept the license for each model before downloading.