LLaMA

LLaMA is Meta's family of open-weight large language models, and the current generation — Llama 4, released April 2025 and still the latest publicly available version as of mid-2026 — marked a genuine architectural shift for the line. Llama 4 Scout and Llama 4 Maverick are the first Llama models built as natively multimodal, processing text and images together without a bolted-on separate vision encoder, and the first built on a mixture-of-experts (MoE) architecture, where only a subset of the model's total parameters activate for any given input rather than the full network running every time. The two models target different deployment scales. Scout is a 17-billion-active-parameter model with 16 experts, small enough to run on a single Nvidia H100 GPU, and stands out for an industry-leading 10-million-token context window — enough to process extremely long documents or codebases in a single pass. Maverick is far larger, roughly 400 billion total parameters across 128 experts, with a still-substantial 1-million-token context window, positioned for more demanding production workloads. A larger model, Behemoth, was announced as still in training at the April 2025 launch and, as of April 2026, has not been publicly released — worth knowing since some older coverage implies it's already available. Who it's for: developers and companies building their own AI products who want an open-weight model they can self-host, fine-tune, or run at the edge rather than relying exclusively on a closed API, and researchers who need to inspect or modify model weights directly. Strengths: genuinely open weights (downloadable and self-hostable, not API-only), free commercial use for most companies under the Llama 4 Community License, native multimodality without a separate vision pipeline, and Scout's unusually large context window relative to its compact, single-GPU-friendly size. Limitations: the license isn't unconditionally open — companies with 700 million or more monthly active users as of April 2025 must request a separate license from Meta rather than using the community license, and multimodal (vision) capabilities are explicitly excluded for EU-domiciled licensees under current license terms, a meaningful restriction for European companies specifically. Behemoth's continued unavailability also means the top of Meta's model lineup isn't yet accessible to anyone. Real-world use cases: a company self-hosting Llama 4 Scout to process very long internal documents entirely on its own infrastructure for data-privacy reasons, a startup fine-tuning Maverick on domain-specific data to build a specialized assistant, and a research team using open weights to study model behavior in ways a closed API doesn't allow. For teams that specifically need open, self-hostable weights rather than an API-only model, Llama 4 remains one of the most capable available options — with the EU vision restriction and license threshold as the two details most worth checking before committing.