Granite 4.1 30B is a dense, decoder-only transformer language model developed by IBM as the flagship of the Granite 4.1 family. Released in April 2026, the model is designed for enterprise-grade applications, offering a balance of high-capacity reasoning and computational efficiency. It was trained from scratch on approximately 15 trillion tokens of diverse, high-quality data using a multi-stage pipeline focused on multilingual instruction following, tool calling, and long-context understanding.
Architecture and Training
The model utilizes a standard dense transformer architecture incorporating Grouped Query Attention (GQA), Rotary Position Embeddings (RoPE), SwiGLU activation functions, and RMSNorm. Its training process involved a five-phase strategy: the initial phases focused on general pre-training, followed by high-quality data annealing and a final stage dedicated to long-context extension. The model underwent rigorous post-training refinement using Supervised Fine-Tuning (SFT) and Reinforcement Learning through Grouped Relative Policy Optimization (GRPO) with Directed Acyclic Policy Optimization (DAPO) to enhance its alignment and specialized task performance.
Key Capabilities
Granite 4.1 30B natively supports 12 languages, including English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. It is optimized for complex enterprise workflows such as Retrieval-Augmented Generation (RAG), summarization, and advanced code generation. Notably, the model features enhanced tool-calling capabilities, allowing it to interface with external APIs and functions using schemas compatible with standard OpenAI function-calling formats. It also supports Fill-In-the-Middle (FIM) operations for specialized coding tasks and agentic tool usage.