Granite 4.2 LLMs: A Serious Look at IBM’s New Reasoning Models

The open-weights AI space moves fast, but IBM's latest release actually deserves your attention. They just dropped Granite 4.2 LLMs, a fresh family of dense, decoder-only models built from the ground up for serious reasoning. Instead of just chasing massive parameter counts, they focused on what smaller teams actually need: practical sizes, real-world utility, and an open license that doesn't tie your hands.

Available in 3B, 8B. 30B variants, these models ditch the standard hype cycle for actual engineering substance. This they emerge from a tough five-phase pre-training strategy across fifteen trillion tokens, extending context windows out to an impressive 512K. Then comes the clever part. Turns out, before throwing the 8B and 30B versions into a multi-stage reinforcement learning gauntlet. IBM layered on supervised fine-tuning for chain-of-thought processing. IBM layered on (interestingly) supervised fine-tuning for chain-of-thought processing.

What makes this pipeline genuinely interesting for builders is the agentic RL block. The these models aren't just reciting text. They learn to interact with live sandboxed environments, writing code, navigating terminals, and calling tools natively using standard formats. Plus, every model includes a thinking switch and a low-effort mode. And the result? That means you can dynamically adjust the reasoning budget depending on whether you're asking a simple syntax question or orchestrating a complex multi-step workflow.

Best of all? They shipped the whole lineup under the Apache 2.0 license and plugged them right into standard inference stacks like vLLM and SGLang. No proprietary lock-in. No absurd pricing tiers. Just solid, well-crafted weights ready to deploy on your own hardware.