IBM Releases Granite 4.2 With Native Reasoning and Agentic RL

IBM has released Granite 4.2, a new family of open large language models designed with native reasoning capabilities and available in 3 billion, 8 billion, and 30 billion parameter sizes. Unlike previous Granite models that functioned as instruction-following assistants, Granite 4.2 is fundamentally built around explicit reasoning processes. Each model in the Granite 4.2 family is capable of generating a chain of thought before providing an answer, and all models feature a "thinking" versus "non-thinking" switch. Additionally, they include a low-effort mode that allocates a limited reasoning budget for simpler queries. These models are architected as decoder-only dense transformers and were pre-trained on approximately 15 trillion tokens. Following pre-training, they underwent a multi-stage reinforcement learning process. For the 8 billion and 30 billion parameter versions, this reinforcement learning chain incorporates an agentic RL block. This block enables the models to learn tasks such as editing code, interacting with a terminal, and executing web searches within secure, sandboxed environments. All three Granite 4.2 language models are distributed under the permissive Apache 2.0 license, meaning they can be freely downloaded, fine-tuned, and utilized for commercial production without licensing restrictions. Alongside the language models, IBM also introduced two Granite Speech 5.0 Turbo CTC models, each with 470 million parameters, to support speech-related applications. The Granite 4.2 models are engineered for broad enterprise deployment. The 3 billion parameter version is suitable for individual developers and startups, capable of running on a laptop using platforms like Ollama or LM Studio, especially with the availability of GGUF quantizations down to Q4_K_M. The 8 billion parameter model is designed for mid-market teams and can operate on a single modern GPU. The 30 billion parameter model is targeted at enterprises requiring significant computational resources, such as those with A100/H100-class hardware, or for FP8/NVFP4 serving using vLLM. Organizations operating under strict regulatory requirements can benefit from the option of on-premises weight deployment. The models are intended for deployment across various industries, including software and developer tooling, financial services, healthcare, telecommunications, the public sector, and contact centers, where the new speech models are particularly relevant. Potential applications include the development of software engineering agents, automation for terminal and DevOps tasks, advanced research and search agents, and robust Retrieval-Augmented Generation (RAG) for long documents, as well as structured tool calling.