IBM released Granite 4.2, its first family of reasoning-focused language models, in three sizes under the permissive Apache 2.0 licence. Each model can switch between a full reasoning mode, a short-budget mode and answering directly, and all support tool calling in the OpenAI function-calling format. Only the two larger sizes went through the additional training that teaches them to act as agents in real coding, terminal and web-search environments.
What changed
Earlier Granite releases were instruction-following assistants without an explicit reasoning mode or agentic training.
What it unlocks
Running an openly licensed model that can switch between deliberate reasoning and direct answers, and drive tools, code edits, a terminal and web search inside agent harnesses.
- 3B, 8B and 30B model sizes
- ~15T pre-training tokens
- 512K-token context extension
What you need to act on it
- self-hosting or a serving stack such as vLLM or SGLang
Sources