Granite 4.1 LLMs: How They’re Built

IBM published a technical walkthrough detailing how Granite 4.1 LLMs were built. The models are dense, decoder-only transformers trained on approximately 15 trillion tokens across five pre-training phases, extending context to 512K tokens. Yousaf Shah and the Granite Team describe supervised fine-tuning on 4.1 million curated samples and a multi-stage reinforcement learning pipeline. The 8B instruct model matches or surpasses the previous Granite 4.0-H-Small (32B-A9B) despite fewer parameters.