IBM's new open models have a thinking dial - and no European exception in the licence

IBM has published Granite 4.2, three reasoning models of 3, 8 and 30 billion parameters, under plain Apache 2.0 with no territorial carve-out - and the middle one resolves 47.67 percent of SWE-bench Verified, a software-repair test that until recently belonged to frontier systems. The design choice that sets the family apart is the thinking dial. Rather than shipping a separate reasoning model, IBM builds the chain of thought into each size and lets the caller pick between full thinking, a low-effort budget and no thinking at all, on a per-query basis. The models write their reasoning inside <think> tags before answering. IBM's own figures put the 30B at 57 on SWE-bench Verified and 89.17 on AIME25, with the 8B close behind at 47.67 and 86.67. The second choice is architectural and goes against the grain: all three are dense. At a moment when nearly every model of this size activates only a fraction of its weights per token, Granite 4.2 runs all of them. That costs compute, but it removes the routing machinery that makes mixture-of-experts models awkward to serve on small hardware. Training was done on rented capacity - an NVIDIA GB200 cluster at CoreWeave - which is a notable detail for a company that sells computers. For a European reader the licence is the point. Over the past week this catalogue has covered several open-weight releases whose agreements exclude the European Union outright or bind only companies above a user threshold. Granite carries Apache 2.0 and nothing else: no territory clause, no acceptable-use annex, not even a licence file with extra conditions in the repository. One caveat belongs next to the headline. IBM lists twelve tested languages and Polish is not among them, and the smallest model behaves very differently from its siblings outside English: on IBM's own multilingual test the 3B scores 27.78 where the 8B reaches 61.06. The shared specification sheet does not warn about that gap. The 3B also skipped the agentic reinforcement-learning stage that the 8B and 30B went through, which is the likelier explanation for why it reports no SWE-bench result at all.