Today we are excited to introduce Granite 4.0 Nano, the smallest models yet in IBM’s Granite 4.0 family. Designed specifically for edge and on-device applications, these compact models deliver strong performance despite their minimal size, reinforcing IBM’s commitment to building powerful, practical AI models that don’t rely on hundreds of billions of parameters. Like all Granite 4.0 models, the Nano variants are released under the Apache 2.0 license and feature native support across widely used runtimes such as vLLM, llama.cpp, and MLX. They were trained using the same advanced methodologies, pipelines, and over 15 trillion tokens of training data that powered the original Granite 4.0 models. This release includes four instruction-tuned models and their corresponding base versions, all built with the efficient hybrid architecture that defines the Granite 4.0 series. The Granite 4.0 Nano models represent a significant leap in capability for sub-billion to approximately one-billion-parameter models—a space where leading developers like Alibaba (Qwen), LiquidAI (LFM), Google (Gemma), and others are actively innovating. When benchmarked across general knowledge, math, coding, and safety domains, Granite 4.0 Nano models consistently outperform similarly sized competitors, demonstrating that high performance is achievable even at a small scale. In addition to general benchmarks, the Nano models show strong results on tasks critical for agentic workflows, including instruction following and tool calling. They achieve top-tier accuracy on the IFEval benchmark and rank highly on Berkley’s Function Calling Leaderboard v3 (BFCLv3), highlighting their suitability for real-world, interactive AI applications. All Granite 4.0 Nano models carry IBM’s ISO 42001 certification for responsible AI development, ensuring they are built and governed according to global standards for ethical and trustworthy AI. This certification adds a layer of confidence for developers and organizations deploying these models. For detailed specifications and performance metrics, including full benchmark breakdowns, refer to the model cards on Hugging Face. As IBM continues to expand the Granite 4.0 family, developers can expect more innovations aimed at making AI more efficient, accessible, and effective—especially for deployment in constrained environments.