IBM releases Granite 4.2 with built-in reasoning modes

Background

A model family can get reasoning ability in two broad ways. One is to train a separate reasoning model and call it when a hard question arrives; the main model stays simple, but two models have to be served and kept in sync, and answers pass through a handoff between them. The other is to train the reasoning behaviour into the main model itself, so a single model produces both the intermediate steps and the final answer.

IBM's Granite family is aimed at enterprise workloads, including code generation and agentic workflows, where a model is expected to call tools and complete multi-step tasks rather than answer a single question ^[1]. The 4.2 generation takes the second route.

The release

IBM released the Granite 4.2 language models on 25 August 2026 ^[1]. Model weights were already available on Hugging Face from 7 August 2026, ahead of the release announcement ^[2]. All sizes in the family are released under the Apache-2.0 licence ^[1].

Reasoning inside the model

The generation adds native reasoning rather than attaching a separate reasoning model to the family ^[1]. In the output stream, the reasoning appears inside think tags, a pair of markers that separates the model's intermediate working from the answer it finally gives ^[1]. Because that working sits in a marked region of the stream, an application can show it, collapse it, or filter it out before the text reaches a user.

Reasoning is not all-or-nothing. Users can switch between full thinking, non-thinking and low-effort modes on a per-query basis ^[1]. The same model can therefore be asked for a long, careful deliberation on one request and a short, direct answer on the next, without loading a different model for each.

Reasoning and tools

The models also reason about which tools to call. IBM presents this as reasoning-augmented tool calling: choosing a tool is treated as a step to be worked out rather than a fixed pattern to be matched ^[1]. For agentic workflows, where a system chains several actions together, the choice of which tool to invoke at each stage is part of what the model has to decide.

Why it matters

Two properties of the release matter for the organisations the family targets. The first is licensing: Apache-2.0 is a permissive open-source licence that allows use, modification and redistribution, subject to its conditions, and it covers every size in the family, so the same terms apply whether a team picks a small model or a large one ^[1]. The second is availability: weights on Hugging Face from 7 August 2026 meant the models could be downloaded and inspected before the release date itself ^[2].

Put together with per-query mode switching, the family gives one set of weights a range of operating points — a full-thinking pass for a difficult coding or planning task, a lighter pass when speed matters more ^[1]. Whether teams use that range to control cost, to keep reasoning out of user-facing output, or to log it for review is left to the deployment.

Granite 4.2 also marks a position in how reasoning is packaged. Rather than shipping reasoning as an add-on model with its own endpoint, IBM made it a property of the main model, exposed through markers in the output stream and a switch the caller controls ^[1].