IBM releases Granite 4.2, a new generation of open-weight AI models for local use
IBM has released Granite 4.2, a new family of open-weight language models that can be downloaded and self-hosted. The release focuses on stronger reasoning capabilities and expanded tool-use skills.
IBM has rolled out Granite 4.2, the latest generation in its family of open-weight large language models built for users who want to download and self-host them rather than rely on cloud services. The new release is available in three sizes: 3 billion, 8 billion, and 30 billion parameters.
As with previous versions, IBM continues to use a decoder-only architecture. All three models natively support a context window of up to 128,000 tokens, allowing them to process significantly longer text inputs in a single pass.
Expanded tool use
The 8B and 30B variants underwent additional training through agentic reinforcement learning, aimed at expanding their capabilities for tasks such as using a terminal, searching the web, or working with external tools. The smaller 3B model also supports tool use, though without the same level of specialized training applied to the larger variants.
A focus on reasoning
According to IBM, Granite 4.2 is the reasoning-focused release within the Granite model family. When researchers describe a model as capable of "reasoning," they are not referring to conscious understanding the way humans experience it. Instead, the term describes functional reasoning, achieved primarily through "chain-of-thought" processing, where the model carries intermediate results forward across multiple steps to reach a conclusion.
The release comes amid growing interest in locally run language models, which give organizations and developers greater control over their data and infrastructure compared to relying on cloud-based AI services.
