IBM releases Granite 4.2 reasoning LLMs with 3B, 8B, 30B sizes and 512K context

Granite 4.2 introduces a family of dense decoder-only reasoning LLMs in 3B, 8B, and 30B sizes. Each model uses Grouped Query Attention with 40 heads, SwiGLU MLP, and RMSNorm, operating in bfloat16 precision. Pre-training from scratch on roughly 15T tokens follows a five-phase strategy that ends with long-context training extending the window to 512K tokens. Supervised fine-tuning mixes agentic and non-agentic data, and post-training adds a multi-stage reinforcement learning pipeline.

Reproducing the models requires the same five-phase pre-training recipe and the supervised fine-tuning data mixture described in the post. The permissive Apache 2.0 license removes legal barriers for commercial deployment. Running the models at their full 512K token context demands compatible inference servers such as vLLM or SGLang.

Only the 8B and 30B variants receive the additional agentic reinforcement learning stage that teaches tool use. The 3B model lacks this stage, so it cannot autonomously invoke tools like code execution or web search. All models include a thinking/non-thinking switch and a low-effort thinking mode, which adds a control layer but does not change the core architecture.

When deployed, the models emit tool calls in the OpenAI function-calling format, allowing direct plug-in to agentic harnesses without extra glue. This compatibility works with existing OpenAI-compatible serving stacks, reducing integration effort. The release under Apache 2.0 also permits modification and redistribution, supporting downstream research and product development.