IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models
MarkTechPost Asif Razzaq ● Covered by 5 sources
IBM just launched Granite 4.2, open reasoning models in 3B, 8B, and 30B sizes. The bigger ones can learn terminal, code, and web-search tasks, not just chat.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
IBM has shipped Granite 4.2, a new set of open language models built for reasoning instead of the usual assistant-style back-and-forth. The lineup comes in 3B, 8B, and 30B parameter versions, and all three are under Apache 2.0, which means they can be downloaded, fine-tuned, and used commercially without a licensing gate.
The main shift is simple: these models can show their reasoning before they answer. They also expose a thinking/non-thinking switch, plus a low-effort mode that uses a smaller reasoning budget on easier prompts. That is a very different posture from earlier Granite releases, which were more about instruction following than explicit deliberation.
IBM trained Granite 4.2 from scratch on roughly 15 trillion tokens. The models are decoder-only dense transformers, with Grouped Query Attention, RoPE, SwiGLU MLPs, RMSNorm, untied embeddings, and bfloat16 precision. The 3B uses 40 layers at 2560 embedding size, the 8B also uses 40 layers but at 4096, and the 30B goes to 64 layers with a 32,768 hidden size in the MLP.
The training recipe is the real story. After pre-training, IBM ran a multi-stage reinforcement learning pipeline. On the 8B and 30B models, that includes an agentic block where the model learns to edit code, use a terminal, and run web searches inside sandboxed environments. The 3B skips that agentic stage and gets foundational RL and alignment only.
IBM’s reported numbers reflect that split. On SWE-Bench Verified, the 30B scores 57.00 and the 8B scores 47.67. On Terminal-Bench 2.1, the 30B reaches 29.24 and the 8B gets 20.56. IBM also released two Granite Speech 5.0 Turbo CTC models at 470 million parameters each, with no LLM backbone, aimed at fast speech-to-text work.
My take — AI-written commentary, not fact-checked reporting
This is the kind of release that makes enterprise AI feel less like a chatbot demo and more like a toolchain. IBM is betting that openness plus real agent training beats splashy closed-model marketing, and that’s a sensible bet. The industry could use more models that know how to work, not just how to talk.
Read more about this at: MarkTechPost