How Are IBM's Granite 4.2 Reasoning Models Built?

Granite 4.2 is IBM’s first stab at dense decoder‑only language models built just for reasoning work. The family comes in three sizes — 3 B, 8 B and 30 B parameters — all sharing the exact same blueprint. That blueprint uses grouped query attention with 40 heads and 8 KV heads, rotary position embeddings with a theta of 10 million, SwiGLU‑activated feed‑forward layers, RMSNorm, and separate input/output embeddings, all done in bfloat16. They start with a context window of 131 072 tokens, but can be stretched to 512 000 tokens in the final pre‑training stage.

Training kicks off from scratch on about fifteen trillion tokens split into five phases. The first two phases soak up broad web‑scale knowledge, while phases three and four tilt the mix toward higher‑quality sources. Phase five then piles on long‑context material to widen the usable window. After that base, supervised fine‑tuning shows the models chain‑of‑thought examples, reasoning traces and agentic trajectories, teaching them to lay out an internal monologue before spitting out an answer.

The next step is a multi‑stage reinforcement‑learning pipeline. Here the 8 B and 30 B models go through an agentic RL block where they learn to call tools, edit and run code, drive a terminal and surf the web inside real sandboxed environments. Every variant — yes, even the 3 B — gets native tool‑calling that spits out OpenAI‑compatible function calls, so they plug straight into agentic harnesses like OpenCode, Pi or OpenHands without any extra glue. A thinking/non‑thinking switch lets users choose how much compute to devote to internal reasoning, and a low‑effort thinking mode offers a middle ground for simpler queries.

All Granite 4.2 checkpoints sit on Hugging Face under the permissive Apache 2.0 license, free for commercial and research use without restriction. You can pull them straight from the Hub, run them with vLLM or SGLang, and toggle the thinking modes to trade latency for depth of reasoning. Because they handle long contexts and tool use out of the box, developers building retrieval‑augmented or agent‑based apps can treat them as a ready‑made foundation for complex, multi‑step workflows.

Source: Hugging Face