IBM Releases Granite 4.2: Open Dense Reasoning Models with Switchable Thinking
On August 25, 2026, IBM officially released Granite 4.2, its first family of dense, decoder-only reasoning language models, available under the permissive Apache 2.0 license on Hugging Face and Ollama.
The new model line ships in 3B, 8B, and 30B parameter sizes, built on a dense all-attention Transformer architecture. Pre-trained from scratch on approximately 15 trillion tokens using a five-phase training pipeline, Granite 4.2 extends its context window up to 512,000 tokens. A standout architectural feature is a switchable native "thinking" mode, allowing developers to toggle step-by-step chain-of-thought deliberation on or off depending on prompt complexity.
For the 8B and 30B checkpoints, IBM applied an additional agentic reinforcement learning (RL) training stage inside sandboxed terminal, code execution, and web-search environments. On vendor-reported benchmarks, IBM claims the 3B model achieves 78.3% on AIME 2025 and 54.8% on GPQA, though independent evaluations from third-party tracking suites like Artificial Analysis remain pending. Unlike mixture-of-experts architectures, the dense design allows predictable hardware provisioning across local enterprise infrastructure.
The release provides enterprise teams with a transparent, self-hostable reasoning alternative to proprietary commercial APIs. By combining Apache 2.0 licensing, OpenAI-compatible tool calling, and native vLLM and SGLang support, Granite 4.2 targets on-premises coding assistants and compliance-critical agent deployments where data retention policies preclude sending internal code to cloud providers.
