IBM has released Granite 4.2, a family of open-source large language models designed with a focus on reasoning and agent capabilities. The release includes three model sizes—3B, 8B, and 30B parameters—all trained on approximately 15 trillion tokens and available under the Apache 2.0 license. The models extend context windows to 512K tokens, enabling them to process significantly longer documents than earlier generations.
The Granite 4.2 family introduces several advanced features aimed at enterprise use cases. Each model includes a thinking/non-thinking mode that allows users to choose how much computational effort to invest in reasoning for a given task, with a low-effort mode for straightforward queries. The larger 8B and 30B models undergo additional agentic reinforcement learning training that enables them to autonomously use tools, write and execute code, navigate terminals, and perform web searches within sandboxed environments. All models support native tool calling through an OpenAI-compatible API.
The release underscores IBM's commitment to open-source AI development and competitive alternatives to proprietary models. By training models specifically for reasoning and agentic workflows, Granite 4.2 targets developers building complex applications that require both analytical depth and autonomous agent capabilities. The Apache 2.0 licensing removes commercial restrictions, positioning the models as viable options for enterprises building custom AI systems.
Key Points
IBM released Granite 4.2, a family of open-source reasoning models in three sizes (3B, 8B, 30B parameters)
Trained on 15 trillion tokens with extended 512K token context window and Apache 2.0 licensing
Flexible reasoning architecture with thinking/non-thinking modes and low-effort reasoning for simpler queries
Larger models trained with agentic RL to autonomously use tools, write code, and search the web
OpenAI-compatible API with native tool calling for seamless enterprise integration