Overview
- IBM publicly released the Granite 4.2 family in 3B, 8B, and 30B parameter sizes as dense, decoder-only models that were trained from scratch and released under the Apache-2.0 license.
- The models were pre-trained on a large 15 trillion-token corpus in five phases that included a long-context phase advertising support up to 512,000 tokens with a released native configuration for 128K tokens and 1 trillion synthetic code tokens from IBM’s CodeAlchemy pipeline.
- The 8B and 30B sizes received an extra agentic reinforcement learning step plus RLHF so they can call tools, edit and run code, use terminals, and search the web for multi-step agent tasks while the 3B supports limited tool calling.
- Independent reporting shows Granite 4.2 is competent but not a benchmark leader, with the 8B often approaching 30B performance and competitors such as Qwen 3.8 outperforming Granite on coding and some other tasks.
- IBM also released two new Granite Speech models and positions the family for enterprise deployment by emphasizing configurable compute-for-reasoning trade-offs and a text-first, modular approach to adding modalities.