Llama

9:06 Llama is Meta's family of AI models whose weights anyone can download and run. The newest are Llama 4 Scout and Llama 4 Maverick, released on Apr 5, 2025: mixture-of-experts models that read text and images. They are free to download from Hugging Face and Kaggle, and free for commercial use under the Llama 4 Community License, except for companies with more than 700 million monthly users. No Llama 5 has followed: since Apr 2026 Meta's new models are the Muse family from Meta Superintelligence Labs, and Meta AI switched from Llama to Muse Spark on Apr 8, 2026.

Pick Llama when a project already uses it, when a tool you rely on is built around it, or for the small Llama 3.2 models (1B and 3B) on phones and laptops. For a new project, newer open models are usually the better start: Gemma, Qwen, DeepSeek, Mistral and gpt-oss all have releases from late 2025 or 2026, and Meta's own Muse Glimmer and Muse Spark Open are newer too. See Run AI Locally.

Quick facts

Llama 4 models

5:55 Source: Meta's Llama 4 announcement and the Scout model card, read Nov 6, 2026. Both released models were trained on data up to Aug 2024 and support 12 languages, including Arabic, Hindi, Spanish and Vietnamese. Meta described Behemoth as a model still in training that served as the teacher for the two smaller ones.

Earlier Llama models

0:26 Llama 3.1 (Jul 23, 2024) came in 8B, 70B and 405B sizes with 128K tokens of context; the 405B was the largest openly available model at the time. Llama 3.2 (Sep 25, 2024) added small 1B and 3B text models for phones and 11B and 90B vision models. Llama 3.3 followed with a 70B text model. Meta still lists the Llama 3 family and Llama 2, each with its own license, on its Llama page.

What does the Llama license allow?

The Llama 4 Community License lets anyone use, change and sell products built on the models, with three conditions in the model card's license text: a company whose products had more than 700 million monthly active users in the previous month must ask Meta for a license; products that ship Llama must show "Built with Llama"; and a model made by training on Llama must start its name with "Llama". Use also falls under Meta's acceptable use policy. Because of these conditions it is not an open source license in the usual sense; Mistral's Small models and Gemma's newer releases use Apache 2.0 and DeepSeek uses MIT.

Reception and the LMArena episode

0:00 Meta claimed Maverick beat GPT-4o and Gemini 2.0 Flash on a broad range of benchmarks. The launch drew criticism when it emerged that the high LMArena score came from an unreleased "experimental" chat version: TechCrunch reported on Apr 11, 2025 that LMArena changed its policies and that the released Maverick ranked below GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro. See AI Leaderboards for how such rankings work.

What came after Llama

Meta Superintelligence Labs' first model, Muse Spark (Apr 8, 2026), is not open weight; Meta said it hopes to open-source future versions. It did release open weights again later in the year: Muse Glimmer (Aug 10, 2026), a 30-billion-parameter model for local agents under Apache 2.0 that Meta says runs on a single consumer GPU, and Muse Spark Open (Oct 15, 2026), with 816 billion parameters (42 billion active) under a Muse Model License with limited commercial terms. Meta's developer site now leads with Muse models and lists Llama under its docs.

How to run Llama locally

The quickest way is Ollama: ollama run llama4 downloads Scout, a 67 GB download, and the Maverick tag is 245 GB, so Llama 4 needs a workstation or server with a lot of memory. On an ordinary laptop, the Llama 3.2 1B and 3B models or Llama 3.1 8B fit far better. LM Studio offers the same models with a desktop interface, and Hugging Face hosts the original weights. See Run AI Locally.

Llama vs the alternatives

• Gemma: Google's open models, from phone sizes up, under Apache 2.0 for the newest; pick it for small, permissively licensed models.
• Qwen: Alibaba's open models in many sizes; a common pick for local coding and multilingual work.
• Mistral: European open models, from Ministral 3 up to the 1-trillion-parameter Mistral Large 4.
• gpt-oss: OpenAI's open-weight reasoning models under Apache 2.0, sized for one GPU or a laptop.
• DeepSeek: large MIT-licensed models with 1-million-token context.

Frequently asked questions

Is Llama free? Yes. Meta's Llama models are free to download and use, including commercially, under the Llama Community License; only companies with more than 700 million monthly users need a separate license from Meta.

What is the latest Llama model? Llama 4 Scout and Llama 4 Maverick, released Apr 5, 2025, are the latest Llama models as of Nov 6, 2026; Meta's newer models are named Muse instead.

Is Llama open source? Llama's weights are open, but its license adds conditions (the 700-million-user limit, "Built with Llama" and naming rules), so it is usually called open weight rather than open source.

Can I run Llama 4 on my computer? Llama 4 Scout is a 67 GB download in Ollama, so it needs a machine with a lot of memory; smaller Llama 3.2 and 3.1 models run on ordinary laptops.

Does Meta AI still use Llama? No. Meta AI switched to Muse Spark, the first model from Meta Superintelligence Labs, on Apr 8, 2026.

Sources

• The Llama 4 herd, Meta AI, Apr 5, 2025
• Llama 4 Scout 17B 16E Instruct model card, Hugging Face, read Nov 6, 2026
• Llama, Meta, read Nov 6, 2026
• Introducing Llama 3.1, Meta AI, Jul 23, 2024
• Llama 3.2: Revolutionizing edge AI and vision, Meta AI, Sep 25, 2024
• Meta's vanilla Maverick AI model ranks below rivals on a popular chat benchmark, TechCrunch, Apr 11, 2025
• Unpack the Latest Llama 4 Herd, Meta Developers, May 19, 2025 (video)
• Meta’s Llama 4 is mindblowing… but did it cheat?, Fireship, Apr 8, 2025 (video)
• Introducing Muse Spark, Meta, Apr 8, 2026
• Introducing Muse Glimmer, Meta, Aug 10, 2026
• Muse Spark Open, Meta, Oct 15, 2026
• llama4, Ollama, read Nov 6, 2026