IBM Unveils Granite 4.1 Model Family

IBM unveiled the Granite 4.1 model family. It updated its enterprise models at once, including language, vision, speech, embedding, and Guardian, and all of them are distributed under the Apache 2.0 license.

The core is Granite 4.1 8B instruct. Although it was designed with a simpler dense decoder-only architecture, it delivered performance similar to or better than Granite 4.0 32B MoE, and it also competed with the latest Gemma and Qwen series on instruction following and tool calling performance with reasoning turned off.

Training used about 15 trillion tokens, and in the later stages it secured up to 512K tokens of context. The RL pipeline was also designed to separately refine instruction adherence, conversation quality, factuality, and mathematical reasoning.

• Granite Vision 4.1 is a document understanding VLM tailored for table, chart, and KVP extraction.
• Granite Speech 4.1 2B recorded a 5.33% WER, and Plus/NAR variants were released together.
• Granite Guardian 4.1 replaces the existing Guardian 3.3 8B and checks safety, quality, and accuracy.
• Granite Embedding Multilingual R2 supports over 200 languages, strengthening multilingual search.

IBM made clear its direction of combining small models in a modular way to run enterprise AI workflows more efficiently.