Granite 4.2 LLMs: How They're Built

Granite 4.2 is the latest family of dense, decoder-only reasoning language models released by the company. This new model family comes in three sizes: 3B, 8B, and 30B. Each model is pre-trained from scratch on a massive 15 trillion tokens using a five-phase training strategy.

The first two phases focus on foundational pre-training, while the next two phases involve mid-training with progressively higher-quality datasets. The final phase introduces long-context training, which extends the context window to 512K tokens.

After pre-training, the models undergo supervised fine-tuning (SFT) to make them reliable instruction-following, reasoning, and tool-using assistants. The SFT data mixture includes both agentic (31.6%) and non-agentic (68.4%) data, totaling around 100 billion tokens. The agentic corpus covers a wide range of domains, including software engineering, tool calling, terminal use, math, search, and action.

To enhance the models' reasoning capabilities, they can produce a chain of thought before their answers and operate in either thinking or non-thinking mode. The thinking mode spends a short reasoning budget on easy questions, while the low-effort mode falls between the two, spending only a minimal reasoning budget on straightforward queries.

The three model sizes share the same architecture and training pipeline but scale differently. All three models are strong reasoners and instruction followers. However, the post-training process shows more significant capabilities. The 8B and 30B models undergo an agentic RL block that teaches them to operate as agents, enabling them to call tools, edit and run code, drive terminals, and search the web inside real environments.

The Granite 4.2 models are served through an OpenAI-compatible endpoint and emit tool calls in the OpenAI function-calling format. They also work seamlessly with agentic harnesses without requiring additional glue. The models are also supported in SGLang, with a ready-to-serve recipe available in the SGLang cookbook.

The technical details of the build process include the decoder-only dense transformer architecture, five-phase training strategy, pre-training on approximately 15 trillion tokens, supervised fine-tuning, and a multi-stage reinforcement learning pipeline. The models are built on a decoder-only dense transformer architecture with specific core components and undergo rigorous quality control before entering the final SFT mixture.