IBM Unleashes Granite 4.2: Its First Open-Weight Reasoning LLM with Agentic RL Training and 512K Context

IBM has released Granite 4.2, its first open-weight reasoning large language model (LLM) family, featuring agentic reinforcement learning (RL) training. This new generation of open Granite models includes 3B, 8B, and 30B parameter variants, all pre-trained on approximately 15 trillion tokens and available under the Apache 2.0 license. For broader context, explore our AI News. For broader context, explore our Top 100 AI Tools.

Granite 4.2: Core Capabilities and Training

The Granite 4.2 models were pre-trained from scratch on approximately 15 trillion tokens. A significant feature across all models is their support for 512K-token context windows, achieved through a five-phase long-context training strategy. This extended context window is designed to enhance the models' ability to process and understand longer sequences of information.

Each model in the Granite 4.2 family also incorporates a unique "thinking / non-thinking" switch and a "low-effort thinking mode." These features aim to provide more control over the model's reasoning processes, potentially optimizing performance for specific tasks.

Agentic RL Training for Advanced Functionality

A key advancement in the Granite 4.2 series is the application of agentic RL training to the 8B and 30B parameter variants. This training was conducted in real sandboxed environments, enabling these models to learn and execute complex tasks such as tool calling, code editing and execution, terminal usage, and web search. This approach positions Granite 4.2 for applications requiring more autonomous and interactive AI capabilities.

The models support native tool calling in the OpenAI function-calling format. Additionally, they are compatible with serving frameworks like vLLM and SGLang, offering flexibility for developers integrating these models into their systems.

Training Data Composition

The supervised fine-tuning (SFT) process for Granite 4.2 utilized approximately 7.2 million samples, totaling around 100 billion tokens. The training data was composed of 31.6% agentic data and 68.4% non-agentic data, reflecting a balanced approach to developing both general language understanding and specialized agentic behaviors.

Why Granite 4.2 Matters

The release of Granite 4.2 signifies IBM's commitment to advancing open-weight LLMs with sophisticated reasoning and agentic capabilities. By providing models with extensive context windows and specialized training for tool interaction, IBM aims to empower developers and enterprises to build more capable and versatile AI applications. The Apache 2.0 license further encourages broad adoption and collaboration within the AI community.

Conclusion

IBM's Granite 4.2 family represents a notable step in open-weight language models, combining deep reasoning capabilities with agentic reinforcement learning. With its large context window, native tool calling, and open-source license, Granite 4.2 offers a foundation for developing advanced AI solutions across various domains. Developers interested in exploring these capabilities can access the models and documentation through the provided resources.

Sources

• Granite 4.2 LLMs: How They're Built
• GitHub - ibm-granite/granite-4.2-language-models · GitHub
• Granite 4.2 Language Models - a ibm-granite Collection
• ibm-granite/granite-4.2-30b
• ibm-granite/granite-4.2-3b · Hugging Face