IBM Releases Granite 4.2 Reasoning Models Under Apache 2.0

IBM has released Granite 4.2, described on Hugging Face as "the reasoning-focused release of the Granite language-model family." The models come in three sizes—3B, 8B, and 30B parameters—and are distributed under the Apache 2.0 license. According to IBM's technical write-up, each model is a dense, decoder-only transformer pre-trained from scratch on roughly 15 trillion tokens using a five-phase strategy, with the final phase extending the context window to 512K tokens. IBM's post states models were then supervised fine-tuned on chain-of-thought, reasoning, and agentic-trajectory data, followed by a multi-stage reinforcement learning pipeline. The 8B and 30B models additionally go through an agentic RL block that teaches them to call tools, edit and run code, drive a terminal, and search the web inside sandboxed environments; the 3B model supports tools without that specialized training. Every model has a thinking/non-thinking switch plus a low-effort mode, and native tool calling in the OpenAI function-calling format. Ars Technica reported the models are designed to be downloaded and self-hosted, citing a 128,000-token context window natively, and framed IBM's pitch as "predictable deployments" rather than raw speed. Ars noted the release lands amid growing interest in local LLMs as cheaper alternatives to frontier cloud models.

• Three sizes: 3B, 8B, and 30B parameters, released under Apache 2.0
• Pre-trained from scratch on ~15 trillion tokens across five phases
• 8B and 30B models trained with agentic RL for tool use, terminal, code, and web search
• Thinking/non-thinking switch, low-effort mode, and native OpenAI-format tool calling