IBM releases Granite 4.2 models for local LLM deployments

IBM has released the latest models in its family of open-weight large language models, designed for download and self-hosting. The newly launched Granite 4.2 family includes variants with 3 billion, 8 billion, and 30 billion parameters.

These models employ a decoder-only architecture. The new releases offer a 128,000-token context window natively, allowing for the processing of extensive text inputs.

The 8B and 30B parameter variants have undergone agent-based reinforcement learning. This enhances their capabilities for tasks such as web searching or using external tools via a terminal. The 3B model also supports tool usage, albeit with less specialized training.

The availability of locally deployable language models addresses growing demand for AI solutions that offer enhanced data security and customization options for businesses.