IBM's new Granite 4.2 models ride the wave of interest in local LLMs
Ars Technica Samuel Axon ● Covered by 5 sources
IBM launched Granite 4.2, new open-weight models you can run yourself. They come with a huge 128,000-token window, and the bigger ones are tuned for tool use.
Based on reporting by Ars Technica, Samuel Axon — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
IBM has shipped the latest versions of its Granite open-weight language models, aimed squarely at people who want to download them and run them on their own systems. The new family is called Granite 4.2, and it arrives in three sizes: 3B, 8B, and 30B parameters.
This is still IBM’s decoder-only approach, so the company isn’t changing the basic shape of the models. What is changing is how much they can hold in working memory. Granite 4.2 natively supports a 128,000-token context window, which gives it a much longer stretch than the short-answer bots most people first met.
The 8B and 30B versions also get an extra round of agent-focused reinforcement learning. IBM says they were trained for expanded abilities such as using a terminal, searching the web, and calling external tools. That puts them closer to the kind of models people use when they want software that can do things, not just talk about things.
The 3B model can use tools too, but it doesn’t get the same specialized training as the larger two. So IBM is drawing a familiar line: the smaller model is there for lighter jobs, while the more capable behavior is reserved for the bigger releases. The timing fits the growing interest in local LLMs, where control, deployment, and self-hosting matter as much as raw fluency.
My take — AI-written commentary, not fact-checked reporting
IBM is reading the room correctly here. The real pitch for open weights now isn’t bragging rights on a leaderboard; it’s control, privacy, and not begging a cloud service for permission to run your own software. The nice part is that IBM is pairing that with actual tool-use training instead of slapping “agent” on the box and calling it a day.
Read more about this at: Ars Technica