OpenModelsHQ

Meta

Menlo Park, USA · Meta Platforms

Meta is based in Menlo Park, USA and is part of Meta Platforms. We track 15 open-weight models from it, and the newest is Llama-Guard-4-12B, released 23 Apr 2025.

Models tracked
15
LLM 13 · Vision 2
Newest release
Llama-Guard-4-12B
23 Apr 2025
Country
USA
Releases timeline
15 Meta releases from 13 Jul 2023 to 23 Apr 2025. The largest is Llama-4-Scout-17B-16E-Instruct at 109B parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
Llama-Guard-4-12BVision23 Apr 202512B
Llama-4-Scout-17B-16E-InstructVision2 Apr 2025109B
Llama-3.3-70B-InstructLLM26 Nov 202470.6B
Llama-3.2-3BLLM18 Sep 20243.2B
Llama-3.2-3B-InstructLLM18 Sep 20243.2B
Llama-3.2-1B-InstructLLM18 Sep 20241.2B
Llama-3.2-1BLLM18 Sep 20241.2B
Llama-3.1-8B-InstructLLM18 Jul 20248B
Llama-3.1-8BLLM14 Jul 20248B
Meta-Llama-3-8BLLM17 Apr 20248B
Meta-Llama-3-8B-InstructLLM17 Apr 20248B
Meta-Llama-3-70BLLM17 Apr 202470.6B
Llama-2-7b-chat-hfLLM13 Jul 20236.7B
Llama-2-7b-hfLLM13 Jul 20236.7B
Llama-2-13b-chat-hfLLM13 Jul 202313B

Origin

Meta’s AI research arm started in 2013 as Facebook AI Research, or FAIR. Yann LeCun directed it until 2018. FAIR released the fastText library in 2016 and PyTorch in 2017. In 2025 Meta Platforms, the Menlo Park, California, parent of Facebook and Instagram, reorganised its AI work under Meta Superintelligence Labs, bringing FAIR, its language-model teams and other groups together.

How it got here

Meta announced Llama on 24 February 2023. The first version’s weights went only to researchers who applied for access, under a non-commercial licence. A torrent of them appeared online on 3 March 2023.

Llama 2 came on 18 July 2023, announced with Microsoft, in 7B, 13B and 70B sizes. This time Meta released the weights for general use, including many commercial uses. Llama 3 came on 18 April 2024 in 8B and 70B sizes, and Llama 3.1, on 23 July 2024, added a 405B model. Llama 3.2 and Llama 3.3 releases sit between Llama 3.1 and Llama 4 on Meta’s Hugging Face page.

Llama 4 arrived on 5 April 2025. Scout and Maverick are mixture-of-experts models that accept text and image input and cover 12 languages. A larger model, Behemoth, was announced but not released. In April 2026 Meta Superintelligence Labs released Muse Spark, which Wikipedia says replaced Llama as the model powering chatbots on Meta’s services.

Licensing pattern

Since Llama 2, each generation has had its own custom licence: broad use is allowed, on terms that differ from Apache 2.0 or MIT. None is an open-source licence in the Open Source Initiative’s sense. The Initiative said Llama’s licences fail its Open Source Definition because they discriminate against some users and fields of use. Meta still calls Llama open source, which Wikipedia notes is disputed.

The licences share several terms. A company whose products had more than 700 million monthly active users on the release date must ask Meta for a licence; the Llama 2, Llama 3 and Llama 4 licences all contain this clause. Each also includes an acceptable use policy. Wikipedia reports that Llama 2’s licence disallowed using its outputs to improve other language models, and that Llama’s licence prohibited military use, which Meta relaxed for the US government and its contractors in November 2024.

The Llama 4 Community License took effect on 5 April 2025. It requires redistributors to keep a notice saying “Llama 4 is licensed under the Llama 4 Community License”.

Read the licence for the specific generation you plan to use.

Release lineage

Newest first
  1. Llama-Guard-4-12B12B
  2. Llama-4-Scout-17B-16E-Instruct109B
  3. Llama-3.3-70B-Instruct70.6B
  4. Llama-3.2-3B3.2B
  5. Llama-3.2-3B-Instruct3.2B
  6. Llama-3.2-1B-Instruct1.2B
  7. Llama-3.2-1B1.2B
  8. Llama-3.1-8B-Instruct8B
  9. Llama-3.1-8B8B
  10. Meta-Llama-3-8B8B
  11. Meta-Llama-3-8B-Instruct8B
  12. Meta-Llama-3-70B70.6B

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.