Llama
Meta's open-weight model family — free to download, self-host, and fine-tune.
Free to download and self-host, Llama anchors the open-weight side of the AI market. The current line is Llama 4 — the natively multimodal, mixture-of-experts Scout and Maverick, with Meta claiming a 10-million-token context window for Scout. The catch is the Llama Community License: not open source by OSI standards, it caps use at 700 million monthly active users, mandates "Built with Llama" attribution, and restricts EU-domiciled companies. The larger Behemoth model was trained but never shipped.
Best for
Teams that need model weights they can host, fine-tune, and audit inside their own infrastructure — regulated enterprises, on-prem deployments, and startups avoiding per-token API dependence.
Llama's edge over closed rivals like GPT-5 and Claude is control, not capability: you own the deployment. Against fellow open-weight players — DeepSeek, Alibaba's Qwen, Mistral — its advantage is ecosystem gravity: nearly every inference stack, fine-tuning tool, and GPU cloud supports Llama first. DeepSeek and Qwen now ship under genuinely permissive MIT/Apache-style licenses, which makes Llama's custom license the weakest link in that comparison.
Core features
Open model access
Download Llama 4 Scout and Maverick weights from llama.com or Hugging Face and run them anywhere — on-prem, VPC, or quantized on a workstation.
Developer ecosystem
First-class support in vLLM, llama.cpp, Ollama, and the major fine-tuning stacks, plus hosted per-token access via AWS Bedrock, Groq, and Together.
Enterprise flexibility
Self-hosting keeps data inside your own boundary — the reason regulated industries shortlist Llama before any closed API.
Pros
- + The most tooling-supported open-weight family — vLLM, llama.cpp, Ollama, and every major GPU cloud target Llama first
- + Free weights eliminate per-token API costs and keep prompts and fine-tuning data inside your own infrastructure
- + Llama 4's mixture-of-experts designs (Scout, Maverick) bring natively multimodal models to self-hostable sizes
Watchouts
- - The Community License isn't OSI open source — the 700M-MAU cap, attribution mandates, and EU restrictions add legal review
- - Frontier momentum is in doubt: Behemoth was delayed indefinitely and Meta has internally debated closing future models
Power List links
People behind Llama
Alternatives
Frequently asked
What is Llama?
Meta's open-weight model family — free to download, self-host, and fine-tune.
Who builds Llama?
Llama is built by Meta, a publicly traded company (META).
How much does Llama cost?
Llama uses open source. Current tiers: Free (open weights). Check the official site before buying — AI pricing changes often.