IBM has shipped Granite 4.2, a family of three reasoner models built from scratch on roughly 15 trillion tokens. The 3B, 8B, and 30B dense decoders share the same architecture—GQA attention with 40 heads, RoPE embeddings with an unusually high theta of 10 million, SwiGLU feed-forward layers—but the interesting story is what happens after pre-training. Every model gets a five-phase training run that stretches context out to 512K tokens, followed by SFT on about 7.2 million samples. That’s roughly 100B tokens of supervised data, and a third of it is agentic.

The agentic SFT corpus leans hard into software engineering, which makes up 69% of it. Tool calling accounts for another 12%, terminal use 8%, with math, search, and generic actions filling out the remainder. IBM generated these trajectories using a grab-bag of harnesses including OpenHands, SWE-agent, OpenResearcher, and Gemini CLI. They also ran LLM-based quality control with GPT-OSS-120B and Gemma 4 acting as judges, filtering out low-scoring or hallucinated samples before anything reached the final mixture.

Then comes the RL pipeline. All three sizes go through multi-stage reinforcement learning, but the 8B and 30B get an additional agentic RL block. This is where the models actually operate inside sandboxed environments—editing code, driving a terminal, searching the web, calling tools—and receive rewards based on how well they perform. That’s a meaningful departure from the usual post-training recipe of preference optimization on static datasets. IBM is essentially running mini-internships for these models inside live environments.

Every Granite 4.2 model ships with a thinking/non-thinking switch, plus a low-effort mode that spends a short reasoning budget on easy questions. Native tool calling comes standard, served through OpenAI-compatible endpoints via vLLM or SGLang. And it’s all Apache 2.0. For enterprises that have been hesitant about proprietary reasoning models or want something they can run on their own infrastructure, that license is the real headline. IBM is betting that reasoning plus agentic competence plus permissive licensing is the combination that finally gets Granite off the bench in production workloads.