LlamaCon 2025 Keynote: Llama 4, Open-Source AI, and AI Future with Zuckerberg & Ghodsi

The Power of Open-Source AI

• 💡 Two years ago, open-source AI was a dream, but now it's a consensus in government and tech, seen as essential for building and deploying AI.
• ✅ Open-source models like Llama can be safer and more secure through auditing, performant at the frontier, and easily customized for specific use cases.
• 🚀 Llama 4 achieved 1.2 billion downloads in just 10 weeks, with thousands of developers creating tens of thousands of derivative models.

Llama 4 Innovations

• 🧠 Llama 4 is the first open-source multimodal model, trained natively on photos and text, making it fast for visual use cases.
• 🌐 It supports 200 languages natively (10x more multilingual tokens than Llama 3) and features a huge context window for entire codebases or documents.
• ⚡ The model focuses on price performance, delivering high intelligence in small packages like Scout (single H100) and Maverick (17B parameters on 8 GPUs).

Building with the Llama API

• 🛠️ The new Llama API allows developers to start building with one line of code, offering speed, ease of use, customization, and no lock-in.
• 🎯 It provides tools for fine-tuning models for specific product use cases, giving users full agency and control over their custom models.
• 📈 The API includes an evaluation area to assess fine-tuned models, allowing users to add graders and dive deep into results.

Real-World Applications & Impact

• 🛰️ Llama 3 is deployed on the International Space Station for astronauts to access documentation without Earth connection.
• 🏥 In healthcare, models help providers like Sophia and Mayo Clinic reduce paperwork and aid diagnosis by processing medical literature.
• 🤝 Databricks customers use Llama for critical applications like Crisis Text Line (detecting self-harm) and FactSet (natural language financial queries).

The Future of AI Development

• 💰 Open-source models like Llama drive pricing pressure, making AI more accessible and unlocking new use cases.
• 🔬 They enable university research and foster a global community that rapidly advances AI through mixing, matching, and distillation.
• 🗣️ Voice interfaces are expected to become a massive modality, with AI becoming more conversational and integrated into daily interactions.
• 🧩 The future will see many specialized AIs, with businesses building custom agents tailored to their specific needs and data.