IBM has released Granite 4.2, an open-source family of decoder-only dense transformer language models available in 3B, 8B, and 30B parameter sizes. Pre-trained on 15 trillion tokens and post-trained using asynchronous Grouped Query Attention Reinforcement Learning (GRPO), the models feature explicit native chain-of-thought reasoning and built-in budget switches. The 8B and 30B variants incorporate a specialized agentic RL block that trains models to edit code, execute terminal commands, and perform web searches in sandboxed environments.
All models ship under the permissive Apache 2.0 license for unrestricted commercial use. Training was conducted on an NVIDIA GB200 NVL72 cluster hosted by CoreWeave using IBM’s CodeAlchemy pipeline for synthetic code generation. Alongside the text models, IBM introduced two 470M-parameter Granite Speech 5.0 Turbo CTC models designed for high-throughput speech-to-text processing without an underlying LLM backbone.
By open-sourcing fully transparent reasoning and agentic models, IBM presents enterprise developers with an alternative to proprietary cloud APIs for complex software engineering and terminal workflows.
Why it matters
Apache 2.0 licensing allows startups and enterprises to self-host advanced agentic reasoning models without per-token API costs.
Integrated agentic RL training provides native tool use and code editing capabilities directly inside smaller 8B and 30B open weights.
Architectural features like variable reasoning switches and speculative decoding offer fine-grained control over inference budgets and latency.
Source: marktechpost.com
