The Release That Changed the Open-Source AI Landscape

When Meta released Llama 1 in February 2023, it was a research model that leaked onto the internet before anyone had fully understood its implications. When Llama 4 launched on April 5, 2025 under the Llama Community License, it was a deliberate, well-publicised release that the developer community had been anticipating for months. The model that once embarrassed the AI safety community has become the most carefully watched open-source launch in the history of machine learning - and ten months on, its impact on the industry is still unfolding.

Meta released two immediately available variants. Llama 4 Maverick has 400 billion total parameters using a mixture-of-experts architecture, activating only 17 billion parameters per inference pass - giving it computational efficiency far beyond what the total parameter count suggests. Llama 4 Scout is the smaller sibling: 109 billion total parameters, 17 billion active, and - crucially - capable of running on a single NVIDIA H100 GPU. That last point matters enormously: a model with GPT-4-class capabilities accessible to individual researchers and small teams without enterprise compute budgets.

What Scout and Maverick Actually Deliver

Both models are natively multimodal - they process text and images together rather than treating vision as an add-on. Scout supports an industry-leading 10 million token context window, enabling document-level reasoning at a scale that no closed model offered at comparable accessibility. Maverick's context window is 1 million tokens, still far beyond most practical requirements. On multilingual tasks, Llama 4's training data diversity is a genuine advantage - the models perform strongly on Hindi, Bengali, Arabic, and other languages where closed models have historically underperformed, making them particularly relevant for the Indian developer ecosystem.

A third model, Behemoth - with approximately 288 billion active parameters and 2 trillion total - was announced but remained in training and limited preview at launch, with no broad public release as of early 2026.

What This Means for the Industry

The AI industry's closed-model revenue model depends on a capability moat. Llama 4 has significantly narrowed that moat. When a self-hostable model performs comparably to paid APIs for many tasks, the value proposition of commercial APIs shifts from raw capability to infrastructure, reliability, safety guarantees, and support. OpenAI, Anthropic, and Google are not standing still - but they are now competing against a free alternative that improves with every community contribution. For Indian developers and startups, the implications are especially significant: running frontier-class AI locally eliminates data sovereignty concerns, removes API costs entirely, and enables use cases that closed API terms prohibit.