Updated
Updated · Ars Technica · Aug 26
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.

Insights

Will the adjustable thinking modes in Granite 4.2 actually prevent the dreaded lost-in-the-middle memory failures?
Can IBM's shift to functional reasoning finally solve the hallucination problem in complex enterprise AI workflows?
How does training an AI to act like a terminal operator change its fundamental understanding of human language?