Granite 4.2 LLMs: How They're Built

IBM has released Granite 4.2, a new family of open-source reasoning models available in 3B, 8B, and 30B parameter sizes. Published under the permissive Apache 2.0 license, these dense, decoder-only models are built from the ground up for complex tasks, featuring native tool-calling capabilities and a "thinking" mode for deliberate problem-solving. The release provides developers with a set of powerful, enterprise-backed models designed specifically for building sophisticated AI agents.

How Granite 4.2 is Built

The Granite 4.2 models are the result of a rigorous, multi-stage training process. Pre-training started from scratch on approximately 15 trillion tokens, using a five-phase strategy that extends the context window to a massive 512K tokens. Following this, the models underwent supervised fine-tuning (SFT) on a corpus of 7.2 million samples rich in chain-of-thought, reasoning, and agentic data. The 8B and 30B models were further refined through an agentic reinforcement learning pipeline where they learned to use tools, write code, and operate a terminal within sandboxed environments.

• Architecture: Decoder-only Transformer
• Attention: Grouped Query Attention (GQA)
• Position Embedding: RoPE with θ = 10,000,000
• Activation: SwiGLU
• Normalization: RMSNorm

Market Implications and Open-Source Strategy

By open-sourcing the Granite 4.2 family, IBM is providing a significant contribution to the development of autonomous AI agents. The models' native support for OpenAI-compatible function calling simplifies integration into existing agent frameworks, lowering adoption barriers for developers. This move positions IBM as a key provider of foundational models for enterprise-grade agentic applications and intensifies competition in the open-source arena, offering a credible alternative to both proprietary systems and other open models that may lack this specific focus on verifiable, tool-using reasoning.

IBM's multi-stage, multi-environment reinforcement learning pipeline, particularly the agentic RL in sandboxed environments, marks a deliberate pivot in open-source model development from generalized instruction-following to specialized, reliable tool-using agents for enterprise workloads.