Introducing the IBM Granite 4.1 family of models

IBM has released the Granite 4.1 model family, its most expansive release to date. The collection includes dense decoder-only language models in 3B, 8B, and 30B sizes, with the 8B instruct model matching or outperforming the previous 4.0 32B MoE model. Key highlights include: Granite Vision 4.1 for document understanding (tables, charts, KVP extraction); Granite Speech 4.1 models achieving 5.33% WER with a novel non-autoregressive variant for higher throughput; Granite Guardian 4.1 for safety moderation and harm detection; and Granite Embedding Multilingual R2 supporting 200+ languages. All models are trained on ~15 trillion tokens with multi-stage RL fine-tuning and released under Apache 2.0. They are optimized for vLLM, SGLang, and llama.cpp and available on watsonx and Hugging Face.