Meta Llama 4

Llama 4 is Meta's fourth open-weight LLM generation, released April 5, 2025 as two mixture-of-experts models. Scout packs 109B total parameters (17B active across 16 experts) with a 10-million-token context window that fits on a single H100; Maverick scales to 400B total (17B active, 128 experts) with a 1M-token window. Both are natively multimodal — a vision encoder trained jointly with the language model — with an August 2024 knowledge cutoff, and both ship under the custom Llama 4 Community License, which permits commercial use for most businesses but is not an OSI open-source license.

The launch was shadowed by a benchmarking controversy: Meta claimed Maverick beat GPT-4o on LMArena, but LMArena later noted the score came from an unreleased experimental chat variant tuned for conversationality, not the public weights — a reminder to treat the vendor's leaderboard claims with care. Meta has since shifted its developer spotlight to the closed Muse line, but Scout and Maverick remain downloadable and widely hosted, with Scout still one of the longest-context freely available models and a durable backbone for RAG-heavy and fine-tuned open stacks.

• Meta-reported: Llama 4 Maverick outperformed GPT-4o on LMArena at launch (provider claim — LMArena later clarified the score came from an experimental chat variant, not the public weights)
• No canonical independent benchmark table verified for the public weights; treat vendor comparisons as provider-reported