Meta Llama
5.6/10Meta's family of open-weight large language models, including Llama 3.x and the multimodal Llama 4, which teams download and run themselves or through third-party hosts.
Data last updated November 2026 · Our methodology
Overview
Meta Llama is a family of open-weight large language models that Meta released between February 2023 and April 2025. The current lineup is Llama 2, Llama 3 (3.1, 3.2 and 3.3) and Llama 4. Llama 4 Scout (109B total parameters, 17B active, 10M-token context) and Llama 4 Maverick (400B total, 1M context) are mixture-of-experts models that accept images as well as text. Organisations download the weights from Hugging Face or Kaggle under Meta's community licenses, then fine-tune and host them on their own hardware or through providers such as Groq, Together AI, Microsoft Foundry and Amazon Bedrock. Meta charges nothing for the weights. Companies whose products exceed 700 million monthly active users need a separate license, and the Llama 4 license excludes companies based in the EU. Your real cost is compute, or per-token fees from whichever host you choose. Llama suits developer and IT teams that need a model they can host privately and customise, with no per-seat fees. The main limitation is that the line has effectively stopped moving forward. In April 2026 Meta switched its own products to the proprietary Muse Spark model and said only that existing Llama models would stay available. It retired the Llama API preview on July 6, 2026. Cloud hosts are also winding Llama down: Google Cloud retired its managed Llama 3.3 API in October 2026, and Bedrock lists Llama models as legacy with an end-of-life date of March 2027. Reviewers had already found Llama 4 behind rivals such as DeepSeek V3 on coding benchmarks at launch.
Our Rating
Who Meta Llama is best for
Best software for DevelopersGood fit
Open weights, broad tooling support and fine-tuning freedom suit engineers building their own AI applications, though the models are aging.
Less suited
IT teams can self-host the models for full data control, but they carry the infrastructure work and get no vendor support.
No license fees and cheap third-party hosting help startups on a budget, but newer open models often perform better.
Self-hosting gives data control, but there is no enterprise support, the 700M MAU clause and EU exclusion apply, and cloud providers are retiring Llama endpoints.
Llama is a raw model with no ready-made app or admin tools, so non-technical small businesses get little from it directly.
Pros & Cons
Pros
- The weights are free to download and use commercially, so self-hosting avoids per-seat and per-token fees from the model vendor.
- Open weights allow full fine-tuning and private deployment, which keeps sensitive data on your own infrastructure.
- The range of sizes, from the 1B and 3B Llama 3.2 models to the 400B Llama 4 Maverick, covers uses from edge devices to multi-GPU servers.
- Llama 4 Scout advertises a 10M-token context window, and both Llama 4 models accept image input.
- Many hosts serve Llama cheaply. Recent third-party trackers list Llama 4 Maverick at around $0.15 to $0.50 per million input tokens on providers such as DeepInfra and Groq.
- Meta publishes companion safety models such as Llama Guard 4 and Prompt Guard for filtering inputs and outputs.
Cons
- Meta moved its own products to the proprietary Muse models in April 2026 and has announced no new Llama releases, so the line is effectively frozen at Llama 4 from April 2025.
- Meta retired the Llama API preview on July 6, 2026, and providers are dropping Llama: Google Cloud retired its managed Llama 3.3 API in October 2026, and Bedrock marks Llama as legacy with end of life on March 30, 2027.
- The Llama 4 license bars companies based in the EU from using the multimodal models, and companies above 700 million monthly active users need Meta's permission.
- Independent testing found Llama 4 Maverick scored 16% on the aider polyglot coding benchmark, behind DeepSeek V3, and critics accused Meta of using a tuned model for its LMArena results.
- G2 reviewers most often complain about limited capabilities in specific use cases, slow code generation and repetitive responses.
- There is no first-party support, admin console or analytics, so teams must build monitoring and governance themselves.
Key Features
Not included
- Image Generation
- API Access
Meta Llama Pricing
Open weights (Llama Community License)
- Commercial use allowed under the community license
- Llama 4 license excludes companies based in the EU
- Derivative distributions must carry Built with Llama branding
- Separate license needed above 700 million monthly active users
- Hosting and compute costs are paid by the user or to a third-party provider
- Download Llama 2, 3.x and 4 weights via Hugging Face or Kaggle after accepting the license
Frequently Asked Questions
Meta has not released a new Llama model since Llama 4 in April 2025. In April 2026 it launched the proprietary Muse Spark and moved its own products to Muse. A Meta spokesperson said current Llama models will stay available as open source but gave no commitment to future Llama versions.
No longer. Meta shut down the Llama API public preview on July 6, 2026 and told developers to move to third-party providers. Meta's current Model API serves Muse models, not Llama.
Meta charges nothing for the weights under its community license. You pay for your own GPUs or for per-token fees from a hosting provider. Recent third-party price trackers list Llama 4 Maverick at around $0.50 input and $0.77 output per million tokens on Groq.
The Llama 4 Community License excludes companies and individuals domiciled in the EU from the multimodal Llama 4 models. Llama 3.x models such as Llama 3.3 70B are still licensed for EU use.
Commercial use is allowed. Organisations whose products exceed 700 million monthly active users must request a separate license, which Meta can grant or refuse. Derivative distributions must show Built with Llama branding. The Open Source Initiative does not consider these licenses open source.
The weights are on Hugging Face under the meta-llama organisation, which lists 70 models, and on Kaggle, once you accept the license. Hosted options include Groq, Together AI, DeepInfra, Microsoft Foundry and Amazon Bedrock, though some clouds are deprecating Llama endpoints.
Meta says Llama 4 Scout can run on a single Nvidia H100 GPU with quantisation. Maverick needs an H100 DGX system or equivalent multi-GPU setup. The smaller Llama 3.2 models run with roughly 4 to 24 GB of memory.
Similar Tools
See all alternatives
Mistral AI's enterprise assistant, renamed Vibe in May 2026. It combines chat, work agents, connectors and coding agents, and can be deployed on-premises or in a private cloud.
Onyx, formerly Danswer, is an open-source AI assistant and enterprise search tool that lets staff search and chat across 50+ connected workplace apps with permissions respected.
CrewAI is an open-source Python framework and enterprise platform (AMP) for building, deploying, and governing multi-agent AI workflows. It is used by developers and enterprise platform teams.
Ranked near Meta Llama in AI Software
See Meta Llama in the listMeta Llama is #320 in AI Software by our score.
