IBM Releases Granite 4.2 Reasoning LLMs with Agentic RL and Tool Calling

IBM has unveiled Granite 4.2, a new family of dense, decoder-only large language models engineered explicitly for advanced reasoning and agentic workflows. Released in 3-billion, 8-billion, and 30-billion parameter configurations under the Apache 2.0 license, the models introduce a dedicated reasoning architecture with built-in chain-of-thought capabilities, flexible thinking modes, and native tool-calling functionality. The foundation of Granite 4.2 rests on a five-phase pre-training strategy executed on approximately 15 trillion tokens. This process progressively shifts from broad web-scale corpora to highly curated, high-quality datasets, culminating in a long-context training phase that extends the sequence window to 512,000 tokens. Supervised fine-tuning follows, utilizing a 7.2-million-sample corpus balanced between agentic and non-agentic domains. To ensure fidelity, IBM implemented rigorous data quality control, employing LLM-based judges, deduplication algorithms, and strict filtering for hallucinations and invalid tool interactions. The models most significant innovation lies in their multi-stage reinforcement learning pipeline. Built on Group Relative Policy Optimization, the training process operates asynchronously to maximize hardware utilization. All three sizes complete foundational reinforcement learning phases focused on verifiable rewards, mathematical reasoning, and skill-specific boosters. The 8B and 30B variants further undergo a specialized agentic training curriculum, teaching them to navigate real-world sandboxes for software engineering, terminal operations, and web research. Every model concludes with reinforcement learning from human feedback to align tone, enforce safety protocols, and optimize reasoning efficiency. Training infrastructure leverages an NVIDIA GB200 NVL72 cluster hosted by CoreWeave, utilizing a non-blocking Fat-Tree NDR InfiniBand fabric for low-latency distributed computing. The software stack is anchored by NeMo-RL and NeMo-Gym, which standardize reward signals and environment interactions across disparate training stages. Independent evaluations demonstrate consistent performance gains across agentic coding, general reasoning, and tool-use benchmarks. The 30B variant achieves state-of-the-art resolve rates on SWE-Bench Pro and Terminal-Bench, while all sizes show marked improvements in long-context retention and multilingual instruction following. To facilitate deployment, IBM has released FP8, NVFP4, MXFP4, and GGUF quantized variants optimized for inference engines like vLLM. The models expose an OpenAI-compatible API, enabling seamless integration with popular agentic frameworks such as OpenCode, Pi, and OpenHands. Granite 4.2 establishes a new benchmark for efficiently scaling reasoning and autonomous tool execution across accessible model sizes.