IBM Launches Granite 4.2 Reasoning Models with Switchable Modes

IBM has introduced Granite 4.2, its latest suite of dense, decoder-only reasoning models spanning 3B, 8B, and 30B parameters. Published under the Apache 2.0 license, the lineup introduces a switchable thinking mode across all sizes, alongside advanced reinforcement learning pipelines tailored for software engineering, terminal execution, and web navigation in the larger variants.

The development pipeline begins with pre-training from scratch on roughly 15 trillion tokens across a five-phase schedule that extends the context window to 512,000 tokens. This foundation is followed by supervised fine-tuning on approximately 7.2 million samples containing chain-of-thought, reasoning, and agentic trajectory data. The core innovation of the release, however, lies in its multi-stage reinforcement learning curriculum, where each phase targets a discrete capability and warm-starts from the previous checkpoint.

All three model sizes undergo foundational reinforcement learning using verifiable rewards, including mathematical problems with clear answers, code evaluated by hidden test suites, and instruction-following tasks equipped with format checkers. Short booster stages are also injected to refine specific operational skills. Meanwhile, the 8B and 30B models progress into an agentic block consisting of three distinct operational environments. In the software engineering stage, managed via the OpenHands harness, models modify live repositories and succeed only when hidden test suites pass. The terminal stage places models inside an interactive shell supporting up to 64 execution turns per rollout, while the search stage requires answering multi-hop queries through live web calls graded by language model judges.

Post-training concludes with standard preference and safety alignment, which incorporates a reasoning-length penalty to mitigate the verbosity often introduced during earlier reinforcement learning phases. Users can access three distinct operating modes through the chat template. The default thinking mode generates a full chain of thought inside designated tags before outputting the final response, whereas the non-thinking mode delivers direct answers. A low-effort setting provides an intermediate option for simpler queries. Additionally, native tool calling is built directly into the template, allowing models to reason about tool selection before execution in the OpenAI function calling format.

Evaluation results released by IBM indicate that performance scales consistently with model size across mathematical, scientific, and coding benchmarks. The reinforcement learning setup utilizes asynchronous group-relative policy optimization, scoring each response against the average reward of other samples generated for the same prompt without relying on a separate value network. Training and generation run on isolated GPU pools to prevent blocking, supported by guardrails that limit policy drift. Infrastructure-wise, the models were trained on an NVIDIA GB200 NVL72 cluster provided by CoreWeave.

Alongside the full precision weights, IBM provides multiple quantized options, including uncalibrated FP8, calibrated NVFP4 and MXFP4, and fourteen GGUF formats designed for reduced memory footprints. The entire collection, along with documentation and source code, is available for public download under the Apache 2.0 license, positioning the suite for developers seeking open reasoning infrastructure.