Llama
Meta's downloadable LLM family, last updated with Llama 4 in April 2025
Overview
Llama is Meta's family of downloadable large language models, from the research-only LLaMA of February 2023 to Llama 4 Scout and Maverick in April 2025. Weights are free under Meta's community licences, which require permission above 700 million monthly users. No Llama model has been released since, Behemoth never shipped, and Meta's newer models carry the Muse name.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Llama
- Self-hosted inference on Llama 4 Scout or Maverick, mixture-of-experts models with text and image input
- Small Llama 3.2 1B and 3B models for on-device and laptop use
- Fine-tunes on private data, with Meta's fine-tuning, distillation and quantisation guides
- Llama Guard 4 and Prompt Guard classifiers to screen prompts and responses
- Air-gapped deployments where no prompt leaves your network
- Commercial use under the Llama community licences, below the 700M monthly-user threshold
Llama 2 and 3 made Meta's family the default starting point for self-hosted and fine-tuned models, and the weights from Llama 2 through Llama 4 remain free to download from Hugging Face and Kaggle. Llama 4 Scout and Maverick, released in April 2025, are the latest; the promised Behemoth never appeared, and the launch was clouded by an LMArena score earned by an unreleased chat variant.
Meta now ships Muse models, including the Apache-2.0 Muse Glimmer, and its hosted Meta Model API offers no Llama. The Llama licence is not OSI open source and carries a 700M-user threshold and EU limits on the multimodal models. Sound for existing deployments; for new open-weight work, compare newer models first.
What we found
Llama is Meta's family of downloadable large language models. The first LLaMA, released in February 2023, was offered to researchers under a non-commercial licence; Llama 2 that July allowed commercial use, and Llama 3, 3.1, 3.2 and 3.3 followed through 2024 in sizes from 1B to 405B parameters. Meta's own Llama page still lists Llama 4, 3 and 2, with weights on Hugging Face and Kaggle.
Llama 4 is the latest and, as of October 2026, the last. Meta released Scout and Maverick on 5 April 2025, mixture-of-experts models with image input and, for Scout, an advertised 10-million-token context window. The same announcement previewed Llama 4 Behemoth, nearly two trillion parameters and "still training". Behemoth weights have never been published: the newest model in Meta's Llama organisation on Hugging Face is a prompt-safety classifier from April 2025. The launch also brought a leaderboard dispute. Meta's blog quoted an LMArena Elo of 1417 for an experimental chat version of Maverick; when the public release build was added, TechCrunch reported it ranking below older rivals, and Neowin counted it 32nd.
Meta's attention has since moved to a new line. In April 2026 Meta Superintelligence Labs introduced Muse Spark, which Meta serves through its own apps and its Meta Model API, whose model list now names Muse and SAM models and no Llama. llama.com redirects to dev.meta.ai, which leads with Muse. Meta's newest open-weight model, Muse Glimmer, released on 10 August 2026, is a 30B model distilled from Muse Spark under Apache 2.0, not a Llama licence. Meta has not announced an end to Llama, and its Llama pages stay up.
The Llama 4 Community License is not open source in the OSI sense: services above 700 million monthly users need Meta's permission, derivatives must display "Built with Llama", and the multimodal rights are not granted to EU-based companies. For self-hosting and fine-tuning the weights remain useful; for new projects, compare Muse Glimmer and other current open models first.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- The Llama 1 model weights were leaked on 4chan within days of being shared with researchers — accidentally accelerating open-source AI adoption by months.
- Meta's Llama 3.1 405B was the first openly available model to match GPT-4 on standard benchmarks, proving that open-weight models could reach frontier performance.
- Llama models power over 25,000 fine-tuned variants on Hugging Face, making it the most forked AI model in history.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Llama's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Llama unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Llama directly →
Spotted something out of date? Suggest an update →
More in LLMs & Chat
