AI BriefingIBM Releases Granite 4.2 Reasoning LLMs
·2026.08.26 00:14
Key point
IBM has released three Granite 4.2 reasoning LLMs featuring 512K context and agentic RL.
Details
IBM has released the Granite 4.2 series of dense (decoder-only) reasoning LLMs under the Apache 2.0 license. Available in three sizes—3B, 8B, and 30B—the models were pre-trained on approximately 15 trillion tokens.
Key technical features include:
- 5-stage pre-training strategy: Data quality was improved stage by stage, with the context window expanded to 512K tokens in the final stage.
- Multi-stage Reinforcement Learning (RL) pipeline: Agentic RL was applied during post-training, enabling the 8B and 30B models to learn capabilities such as tool calling, code execution, terminal manipulation, and web search in real sandbox environments.
- Reasoning mode switching: All models support 'thinking' and 'non-thinking' modes, as well as a 'low-effort' mode that uses a short reasoning budget for simple questions.
- Native tool calling: Tool calling is supported via OpenAI-compatible endpoints (such as vLLM) using the OpenAI function calling format, facilitating easy integration with agent frameworks.
The SFT data included 31.6% agent-related data, leveraging data generated from various agent harnesses such as OpenHands and SWE-agent.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.