IBM Granite 4.2: Apache 2.0 agents you can run yourself

IBM released Granite 4.2 on August 25 — three open-weight reasoning models under Apache 2.0, sized to run on hardware you can actually buy. That license and those sizes matter more than the benchmark table, because open-weight AI agents change the unit economics of every boring, repetitive job in your business from cents-per-call to electricity.

What actually happened

Per IBM Research, Granite 4.2 comes in 3B, 8B and 30B parameter sizes, all released under Apache 2.0 — download, modify, deploy, no royalty. IBM's technical writeup on Hugging Face fills in the specifics: dense transformers with grouped query attention, a native context window of 131,072 tokens, pretraining on roughly 15 trillion tokens, and three inference modes — thinking, non-thinking, and a low-effort mode that spends reasoning budget only on hard questions, toggled by an enable_thinking flag.

The part that matters for automation is the training. The 8B and 30B models went through an agentic RL stage in real sandboxed environments: calling tools, editing and running code, driving a terminal, searching the web. That shows up in the numbers. The 30B scores 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1; the 8B scores 47.67 and 20.56. On knowledge, the 30B hits 77.60 on MMLU-Pro and 89.17 on AIME25.

Distribution is broad on day one — Hugging Face, Ollama, LM Studio, OpenRouter, Replicate, CoreWeave and watsonx among others. So you can test it tonight on a laptop before deciding anything.

Why local inference matters for your business

Nobody should read 57% on SWE-Bench and conclude Granite replaces a frontier model on your hardest work. That is not the trade. The trade is which jobs are boring enough to run for free.

Look at your token bill and sort by volume, not by difficulty. Classifying support tickets. Extracting fields from a PDF invoice. Normalizing product descriptions. Drafting the first version of a reply a human will edit anyway. Those are high-count, low-stakes, and they are exactly what a 8B model with a 128K window handles on a machine that costs less than four months of API spend. Route the hard reasoning to a frontier API; stop metering the easy stuff.

Apache 2.0 is the other half. A permissive license means the weights you tested are the weights you keep — no deprecation notice, no retired model alias, no pricing change in November. For a workflow that has to behave the same way in two years, that stability is worth more than a few benchmark points.

Run the comparison on your own data before you move anything. Same twenty tasks, frontier model versus Granite 8B, and count the failures that would actually reach a customer.

Sources: IBM Research, Hugging Face.