Granite 4.1 3B

Granite 4.1 3B is a compact, dense decoder-only language model developed by IBM and released in April 2026 as part of the Granite 4.1 family. Designed for enterprise-grade applications, it offers a balance of efficiency and performance suitable for local deployment, edge computing, and high-volume workloads. The model was trained from scratch on approximately 15 trillion tokens using a multi-stage strategy that prioritizes data quality and staged refinement. The architecture is built on a dense transformer framework incorporating Grouped Query Attention (GQA), Rotary Position Embeddings (RoPE), and RMSNorm. It utilizes SwiGLU activations and shared input/output embeddings to optimize memory usage. A defining feature of the Granite 4.1 series is its long-context capability, which supports windows of up to 512,000 tokens, enabling the processing of extensive documents and multi-turn conversations. Granite 4.1 3B is fine-tuned for tasks including instruction following, tool calling, and multilingual dialogue. It natively supports 12 languages, including English, German, Spanish, French, Japanese, and Chinese. In comparison to previous versions, the 4.1 models demonstrate improved capabilities in mathematical reasoning, retrieval-augmented generation (RAG), and coding tasks such as Fill-In-the-Middle (FIM) completions. Released under the Apache 2.0 license, the model is intended to serve as a foundation for AI assistants and agentic workflows. IBM encourages the use of the model alongside Granite Guardian for safety monitoring and risk detection in production environments, ensuring alignment with enterprise safety standards.