IBM’s Granite 4.2 family includes dense reasoning models with 3 billion, 8 billion and 30 billion parameters, released under Apache 2.0. All three support native tool calling and a choice of thinking, non-thinking or shorter, low-effort reasoning.
The larger two models receive an additional agent-training sequence. IBM describes reinforcement-learning stages in which they edit repositories, run terminal commands and search the web inside working environments. The 3B model follows a shorter training path without that block; it still supports tool calls.
Software-engineering tasks are rewarded against hidden tests. Terminal tasks use completion checks, while search tasks rely on a model judge because answer quality is less easily reduced to a fixed verifier.
IBM provides model weights, serving guidance and examples for connecting OpenAI-compatible endpoints to coding tools.