IBM releases Granite 4.2 with native reasoning and agentic RL for open enterprise models

IBM has released Granite 4.2, a family of open-source reasoning language models in 3B, 8B, and 30B parameter sizes, built around explicit chain-of-thought capabilities rather than simple instruction following. The models are decoder-only dense transformers pre-trained on roughly 15 trillion tokens and post-trained via a multi-stage reinforcement learning chain that includes agentic RL for the larger variants.

• Granite 4.2 exposes a thinking/non-thinking switch and a low-effort mode, with the 8B and 30B models trained using agentic RL to edit code, drive terminals, and run web searches in sandboxed environments.
• The 30B model achieves 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1, while the entire family is licensed under Apache 2.0 for commercial use.
• IBM also released Granite Speech 5.0 Turbo CTC, a 470M-parameter model using connectionist temporal classification that reports an RTFx throughput near 12,600 on a single H200.

The release provides enterprises with deployable, open-weight models capable of complex reasoning and agentic tasks, alongside high-throughput speech transcription tools for regulated industries.