Bottom line: Meta Llama is Meta’s family of free-to-download open-weight language models, and Llama 4 Scout and Maverick from April 2025 are still the newest release – Meta has shipped no Llama since, moving its frontier work to the closed Muse Spark line before returning to open weights under the Muse brand in August 2026. Llama remains a credible free option for self-hosting, but it is no longer where Meta’s newest models land.
| What it is | Meta’s open-weight LLM family. The current top models are Llama 4 Scout (17B active of 109B total, 10M-token context) and Llama 4 Maverick (17B active of 400B total), both mixture-of-experts and natively multimodal. |
| Best for | Self-hosting and on-premises deployment, fine-tuning on your own data, and any workload where you need the weights themselves rather than an API you cannot inspect or pin. |
| Weakest at | Being current. No new Llama since April 2025, Llama 4 Behemoth was previewed but never shipped, Meta retired the Llama API public preview on 6 July 2026, and llama.com now redirects to Meta’s developer site. |
| Pricing | Weights are free to download from Hugging Face and Meta under the Llama Community Licence, which is not OSI-approved and adds conditions for very large deployments. Hosted access is sold per token by third parties such as AWS Bedrock, Groq and Together. |
| Our take | Still worth running if you need open weights today and Llama 4 clears your quality bar. If you are choosing a new open-weight base in late 2026, benchmark it against Qwen, DeepSeek and Meta’s own Muse Glimmer before committing. |
Independent review. No affiliate links. Last checked 23 August 2026.
Meta Llama (Large Language Model Meta AI) is a family of advanced AI models developed by Meta (formerly Facebook) for natural language processing tasks. These models, including Llama 2 and the more recent Llama 3, are designed to be efficient, scalable, and open-weight alternatives to proprietary AI models. They excel in text generation, summarization, and reasoning tasks, making them useful for developers, researchers, and businesses. Meta has positioned Llama as a competitor to OpenAI’s GPT models, emphasizing transparency and accessibility in AI development.
Frequently asked questions
What is Meta Llama?
Llama is Meta’s family of open-weight large language models, meaning the trained weights can be downloaded and run on your own hardware rather than only called through an API. The current generation is Llama 4, released in April 2025 as Scout and Maverick, both mixture-of-experts models with native multimodal input.
Is there a Llama 5?
No. As of August 2026 the newest Llama release is still Llama 4 Scout and Maverick from April 2025. Meta’s frontier work moved to its closed Muse Spark line under Meta Superintelligence Labs, and in August 2026 Meta returned to open weights under the Muse brand rather than releasing a Llama 5.
Is Meta Llama free to use?
The weights are free to download and run, including for commercial use, with no per-token charge if you host the model yourself. You still pay for the hardware or cloud instance. If you would rather not self-host, providers such as AWS Bedrock, Groq and Together sell hosted Llama access priced per token.
Is Llama actually open source?
Not by the strict definition. Llama is released under the Llama Community Licence, which is not an OSI-approved open source licence. It permits broad commercial use but adds conditions, including a threshold above which very large deployments need separate permission from Meta. The accurate term is open weights rather than open source.
What happened to Llama 4 Behemoth?
Behemoth was previewed in April 2025 as a roughly two-trillion-parameter teacher model used to distil Scout and Maverick, but it has never been released publicly. Meta has not formally cancelled it, and more than a year on it remains unshipped. Treat it as shelved rather than forthcoming when planning around it.
Is the Llama API being shut down?
Meta retired the Llama API public preview on 6 July 2026, and llama.com now redirects to Meta’s developer site, which promotes the Muse line instead. The API’s retirement does not remove the models themselves; Llama 3 and Llama 4 weights remain downloadable and are still served by multiple third-party inference providers.
Where can I download Llama models?
Llama 3 and Llama 4 weights are published on Hugging Face under the meta-llama organisation and via Meta’s own downloads page, both requiring you to accept the Llama Community Licence first. For local use without manual setup, runners such as Ollama and LM Studio pull quantised Llama builds directly.