IBM's new Granite 4.2 models ride the wave of interest in local LLMs

IBM released Granite 4.2, an open-weight decoder-only LLM family available in 3 billion, 8 billion and 30 billion parameter sizes, each supporting a native 128 k-token context window. The 8 B and 30 B models incorporate an agentic reinforcement-learning stage that enables terminal access, web search and external-tool usage, while the 3 B variant can use tools but lacks the specialized training of the larger models.