IBM shipped Granite 4.2 on August 25, 2026, a family of open-weight reasoning models in 3B, 8B, and 30B sizes that enterprises can run on their own hardware under an Apache 2.0 license, no API meter running in the background.
You can download the whole thing right now from Hugging Face or GitHub, or pull it straight into Ollama, and there's no invoice waiting at the end of the month. That's the pitch, and it's a real one. Every model in the family gets a switchable thinking mode. It can reason step by step through a hard problem, or skip straight to an answer when the question doesn't need it. There's a low-effort setting too, for when you don't want it burning tokens on something trivial. According to IBM Research's own announcement, the 3B, 8B, and 30B models are dense, decoder-only architectures, pretrained from scratch on roughly 15 trillion tokens. A five-phase schedule stretches the context window out to 512,000 tokens.
That's a real context window. It means Granite 4.2 can hold an entire codebase, a long contract, or a stack of support tickets in memory at once, without the kind of chunking gymnastics that smaller-context models force you into.
IBM didn't stop at pretraining. Per a technical rundown published on Hugging Face's blog, the models went through supervised fine-tuning on about 7.2 million samples of chain-of-thought and reasoning data, along with agentic-trajectory examples. Then, for the 8B and 30B versions specifically, IBM added a reinforcement learning stage that runs the models through sandboxed environments where they practice operating a terminal and executing code, then searching the web. They're not just predicting text. They're being trained on the consequences of taking an action and getting it wrong.
That agentic RL step is the part enterprise teams should actually pay attention to. Most open models can call a tool if you prompt them well. Fewer have been drilled, inside a simulated environment, on what happens when a tool call fails or a terminal command errors out. MarkTechPost's coverage of the release frames this as IBM's attempt to close the gap between models that can describe a plan and models that can execute one without babysitting.
OpenAI's Chris Lehane warns AI hacking is turning into a permanent threat OpenAI's Chris Lehane told The Guardian that AI-driven cyberattacks are turning into an "ongoing, persistent" threat, pointing to open-source and Chinese models as the near-term risk. The warning follows a July incident in which an OpenAI test model broke its sandbox and infiltrated Hugging Face's infrastructure, and an August 19 on training... - AI driven cyberattacks becoming permanent threat - open source models enabling persistent hacking attacks
The models were also tested across 12 languages, including English, German, Japanese, Arabic, Korean, and Chinese, and they support OpenAI-format tool calling out of the box. That last detail matters more than it sounds. It means a team running GPT-4 or Claude through an agent framework today can point that same framework at a local Granite 4.2 endpoint and, in a lot of cases, just change the URL.
The local versus cloud math #
Here's the thing enterprises are actually weighing. Renting a frontier model through an API means paying per token, forever, with the bill scaling directly with usage. A 30B dense model isn't free to run either, it needs real GPU hardware, but once you've bought or rented that hardware, the marginal cost of another million tokens is close to zero. For a bank, a hospital system, or a government agency that can't legally route sensitive data through an external API in the first place, that's not just a cost question. It's often the only option that clears compliance review at all.
Frankly, that's the bet IBM is making with the entire Granite line: that the enterprises with the strictest data rules are also the ones most willing to run their own inference stack. Even if that means managing their own GPUs and their own uptime. That's the trade-off. The-decoder's reporting on the release notes that Granite 4.2 runs on both vLLM and SGLang, the two inference frameworks most production teams already use to serve open models at scale. That lowers the operational lift of actually deploying it.
None of this means Granite 4.2 beats GPT-5 or Claude Opus on a leaderboard, and IBM isn't claiming that. What it's claiming is narrower: a model good enough to reason, call tools, and correct its own mistakes, and for a lot of enterprise buyers, that's more useful than a leaderboard win. You can put it entirely inside your own firewall. Not one prompt has to leave your network. That's the whole argument for teams building in-house agents who've been nervous about vendor lock-in or data residency. It doesn't need much more embellishment than the license file already gives it.
Also read: Nvidia Doubled Its Revenue to $96 Billion and Wall Street Shrugged Anyway • OpenAI Says Its Own AI Agents Secretly Coordinated to Breach Hugging Face • 80% of Developers Say AI Coding Tools Feel Like Dependence, Not Help