IBM's Granite 4.2 targets the local LLM wave with reasoning and tool use

IBM has released Granite 4.2, the latest version of its open-weight large language model family, designed for developers who want to download and self-host rather than pay per token to a cloud API. The lineup includes 3B, 8B, and 30B parameter variants, all with a 128,000-token context window, and the 8B and 30B models receive additional training for agentic tasks like using a terminal, searching the web, and calling external tools. The 3B model supports tool use too, but with less specialized training.

The headline feature is a shift toward reasoning. IBM calls Granite 4.2 "the reasoning-focused release of the Granite language-model family," meaning the models are trained to engage in chain-of-thought and carry intermediate results through multiple steps. That yields more rigorous and accurate answers in many cases, but also slower response times and higher compute overhead—a trade-off developers will need to weigh when deploying locally.

Why local is having a moment

The launch rides a wave of interest in local LLMs, driven by the escalating cost and compute crunch of frontier cloud models from Anthropic, OpenAI, and others. For many organizations, running smaller open-weight models on their own infrastructure is becoming a practical alternative, especially for tasks that don't require frontier-level intelligence. Granite's pitch here is predictability: these models are not the fastest or most innovative in benchmarks, but they offer stable, repeatable behavior for enterprise deployment.

For developers, Granite 4.2 fits into the growing ecosystem of model routers—AI tools that send each prompt to a model matched to the task's complexity, balancing speed, cost, and capability. A small router could dispatch a simple question to the 3B model and only escalate to the 30B (or a frontier API) when the task demands deeper reasoning.

The models are also a draw for hobbyists and researchers who want to experiment on local hardware without per-token fees. With sizes from 3B to 30B, there's an option for everything from a laptop demo to a production server, all under an open-weight license that invites tinkering and fine-tuning.