Meta Releases Llama 4 with Scout and Maverick Models, Natively Multimodal and 10M Token Context

Meta launched Llama 4 Scout and Llama 4 Maverick on April 5, 2026, framing the release as a direct response to competitive pressure from Chinese open-source models like DeepSeek. Both models support context windows up to 10 million tokens and are natively multimodal — capable of processing text, images, and documents in a single unified architecture. Maverick, with 17 billion active parameters across 128 experts, is positioned as a direct competitor to GPT-4o and Gemini 2.0, with strong benchmark results across coding, reasoning, and multilingual tasks.

The most significant third model, Llama 4 Behemoth, remains in training but has already generated significant industry attention. Early benchmark data shared by Meta indicates Behemoth outperforms GPT-4.5 and Claude Sonnet 3.7 on STEM-focused evaluations, though Meta has not released full third-party benchmark results. The company stated that Behemoth represents a major step up in capability from the Scout and Maverick models and will be released later in 2026 once safety evaluations are complete.

Alongside the model releases, Meta announced a shift toward enterprise licensing tiers for the Llama API — a notable commercial pivot for a model family that had previously been distributed as fully open-weight. Analysts at TechTarget noted that the move mirrors the licensing evolution of other major open-source AI projects, as Meta seeks to monetize its AI infrastructure investments while maintaining the open-weight distribution that has made Llama models the most downloaded AI models globally. The Scout and Maverick weights remain freely downloadable under Meta's community license.