IBM releases Granite 4.2 models for running AI on your own servers

In short: IBM has released Granite 4.2, a set of AI language models that organizations can download and run on their own machines, with a focus on more step by step problem solving.

What happened

IBM rolled out new versions of its Granite “open-weight” language models. “Open-weight” means the model’s core files are available to download, so a company can host it themselves instead of sending requests to a third party.

Granite 4.2 comes in three sizes: 3B, 8B, and 30B. The “B” stands for billions of parameters, which are like the model’s internal settings. In general, bigger models can handle tougher tasks, but they also need more computing power.

IBM says Granite 4.2 is its “reasoning-focused” release. In AI, “reasoning” usually means the model is better at working through a problem in multiple steps and keeping track of intermediate results (like showing its work in math). IBM notes this can improve accuracy for some tasks, but it can also make responses slower and more expensive to run.

Two of the models, the 8B and 30B versions, also received extra training aimed at “agentic” use. That means they are designed to do more than chat, such as using tools, searching the web, or interacting with a computer terminal (a text based control panel). The 3B version can use tools too, but without the same specialized training.

Why it matters

Many companies are looking for alternatives to paid, cloud hosted AI because of cost, speed, and data control concerns. Models like Granite are attractive because they can run locally, which can make deployments more predictable and can help keep sensitive information inside the organization.

Source: Arstechnica