IBM Rolls Out Granite 4.2 in 3B, 8B and 30B Sizes With 128,000-Token Context
Updated
Updated · Ars Technica · Aug 26
IBM Rolls Out Granite 4.2 in 3B, 8B and 30B Sizes With 128,000-Token Context
3 articles · Updated · Ars Technica · Aug 26
Summary
IBM released Granite 4.2 as an open-weight, self-hosted LLM family, positioning it as the reasoning-focused version of its Granite line.
Three variants—3B, 8B and 30B parameters—use a decoder-only architecture and natively support a 128,000-token context window.
The 8B and 30B models also undergo agentic reinforcement learning aimed at multi-step tool use, including terminal access, web search and external tools.
IBM said the 3B model supports tools as well, but without the same specialized training, underscoring the lineup's push toward functional reasoning rather than human-like understanding.