Granite 3.2 2B Instruct

Granite 3.2 2B Instruct is a 2-billion-parameter open-weight language model from IBM's Granite team, released on February 26, 2025. Built on top of Granite 3.1 2B Instruct and trained with a mix of permissively licensed open-source data and internally generated synthetic data, it was IBM's first small model family to ship experimental chain-of-thought reasoning capabilities.

Its signature feature is controllable reasoning: the thinking mode can be enabled or disabled at inference time, so enterprises only pay the compute cost of reasoning when a task actually needs it. IBM applied Thought Preference Optimization, a reinforcement learning technique, to keep reasoning strong without degrading general performance. The model covers summarization, classification, extraction, QA, RAG, code tasks, function calling, and multilingual dialog across 12 languages.

Weights are released under the Apache 2.0 license on Hugging Face and the model is also available through IBM's watsonx.ai platform. The 8B sibling shipped the same day for users who want more capacity in the same package.