Meta Llama

Llama is Meta's family of open-weight language models — you download the weights, run them on hardware you control, and pay nothing per query. The current generation is Llama 4, built on a mixture-of-experts architecture: Scout activates roughly 17B parameters per token out of about 109B total, Maverick activates a similar 17B out of roughly 400B. Both handle text and images natively rather than through a bolted-on vision adapter. A third and much larger model, Behemoth, was previewed at launch but has not shipped as a weights release.

The competitive picture has changed sharply since the entry was first written. Llama 4's April 2025 launch drew criticism on two fronts: the version that posted a headline LMArena score was an experimental chat variant rather than the released weights, and independent long-context evaluations found retrieval quality degrading far below the advertised 10M-token ceiling. In the same window, open-weight releases from Chinese labs — DeepSeek, Qwen, and others — matched or passed Llama on most public reasoning and coding benchmarks. Llama is no longer the default answer to "what's the best open model," and pretending otherwise would be dishonest.

What it still owns is distribution. No other open model runs in as many places: llama.cpp and Ollama for local inference, vLLM and TensorRT-LLM for serving, thousands of fine-tunes on Hugging Face, and managed endpoints on AWS Bedrock, Azure AI Foundry, Google Cloud, Together, Groq, and Fireworks simultaneously. Meta also operates its own Llama API and a consumer Meta AI assistant across WhatsApp, Instagram, and Messenger. One caveat worth knowing before you build: the Llama license is not OSI open source. It caps use by companies above 700M monthly active users, and the Llama 4 multimodal weights are explicitly not licensed to entities domiciled in the EU — a restriction that has surprised more than one European team mid-project.