Granite 4.2 ships a self-hostable reasoning tier under Apache 2.0, and its data says coding agent

The fine-tuning mixture repays a few minutes with a calculator. The agentic block comes to roughly 2.3 million samples ^[16], and software engineering accounts for 69 percent of it ^[9], which puts SWE trajectories at about 22 percent of everything the models were fine-tuned on ^[13]. The rest of that block is thinner: tool calling at 12.1 percent, terminal use at 8.0 percent, then math, search and action ^[9]. The listed slices only add to 93.6 percent, so part of the agentic corpus is unaccounted for in the post ^[21].

The non-agentic figures need a reading before they mean anything. IBM lists instruction following 18.8, coding 18.8, math 14.6, multilingual 7.0, science 5.4, reasoning 3.0 and safety 0.8 ^[10], and those add to exactly 68.4, the same number given for the non-agentic share of the whole corpus ^[17]. They are shares of the full mixture, then, not of their own block. Under that reading the category actually labelled reasoning is 3 percent of the mixture, about a seventh of the software engineering volume ^[14], while SWE trajectories plus non-agentic coding come to roughly 41 percent ^[15]. IBM describes 4.2 as the release that adds explicit reasoning to a family that was previously instruction-following assistants ^[7]. By data volume it is a code and terminal model that can also think out loud.

Two more numbers from the same paragraph of the build post. About 35 of the 100 billion SFT tokens are not trainable ^[18], so roughly a third of that corpus is context the gradient never touches. And the entire supervised stage is under 0.7 percent of the pre-training token budget ^[19], which is a thin coat over a conventionally trained base.

On serving, the attention layout is 40 query heads to 8 KV heads ^[11], a five to one ratio ^[20]. That ratio is what makes the long-context phase ^[2] tractable to serve: cache is one fifth of what an equivalent multi-head configuration would want at the same sequence length. RoPE theta is set to 10 million, weights are bfloat16, input and output embeddings are untied ^[11]. Nothing unusual in the stack, which is rather the point, because at long sequence lengths the capacity question is cache rather than parameters.

The agentic trajectories were generated across a named set of scaffolds, among them OpenHands, SWE-agent, Terminus-2, Codex and Goose, mixing open datasets with IBM's own synthetic RL environments ^[12]. That breadth is also a distribution: a model trained to call tools inside those harnesses has learned their conventions. IBM's integration claim is that an OpenAI-compatible endpoint under vLLM emits function calls in OpenAI format and plugs into agentic harnesses without extra glue, with SGLang supported as well ^[6]. That is a claim to test against your own harness rather than accept.

What the material does not contain is a scoreboard. The post walks through architecture, data and the RL stages, and there are no evaluation figures here to place the 30B against anything else. Apache 2.0 across all three sizes ^[3] is what earns the family an evaluation slot at all; the mixture is what tells you which evaluation to run first.