IBM has unveiled Granite 4.2, a new family of open reasoning language models available in 3B, 8B, and 30B parameter sizes. Unlike earlier Granite iterations that functioned primarily as instruction-following assistants, Granite 4.2 is architected around explicit reasoning capabilities. Each model can generate a chain of thought before providing an answer, and all versions feature a thinking/non-thinking toggle plus a low-effort mode that allocates a minimal reasoning budget for straightforward queries.
The models are decoder-only dense transformers, pre-trained from scratch on approximately 15 trillion tokens, followed by a multi-stage reinforcement learning (RL) post-training pipeline. For the 8B and 30B variants, this pipeline includes an agentic RL component where the model learns to edit code, operate a terminal, and perform web searches within real sandboxed environments. All three models are released under the permissive Apache 2.0 license.
In parallel, IBM also introduced two Granite Speech 5.0 Turbo CTC models with 470M parameters each, designed for efficient speech processing tasks. This release marks a significant step toward bringing advanced reasoning and agentic capabilities to open enterprise AI, aligning with the broader 2026 trend of making sophisticated AI functionalities more accessible and transparent for business applications.
via MarkTechPost