IBM introduces granite 4.2 with enhanced reasoning and reinforcement learning capabilities

IBM has launched Granite 4.2, a new series of open-source reasoning language models available in 3B, 8B, and 30B parameter configurations. Unlike previous versions, which were primarily instruction-following assistants, Granite 4.2 is designed with explicit reasoning at its core. Each model can generate a chain of thought before providing an answer and includes a thinking/non-thinking switch along with a low-effort mode that applies limited reasoning to simpler queries. The models are decoder-only dense transformers, pre-trained on approximately 15 trillion tokens and further refined through a multi-stage reinforcement learning process. The 8B and 30B versions feature an agentic reinforcement learning block that enables the models to perform tasks such as code editing, terminal navigation, and web searches within secure sandboxed environments. All three models are released under the Apache 2.0 license, allowing for free download, fine-tuning, and commercial use without licensing restrictions. IBM also introduced two 470M-parameter Granite Speech 5.0 Turbo CTC models, which use connectionist temporal classification to convert audio into text. These models achieve an RTFx throughput of nearly 12,600 on a single H200 GPU, surpassing current leaders on the Open ASR leaderboard. The architecture of Granite 4.2 includes Grouped Query Attention with 8 KV heads, RoPE with θ = 10,000,000, SwiGLU MLPs, RMSNorm, untied input/output embeddings, and bfloat16 precision. The 3B model uses 40 layers with an embedding size of 2560, while the 8B model employs 40 layers at 4096 and the 30B model uses 64 layers with an MLP hidden size of 32,768. The published architecture supports a sequence length of 131,072 tokens, with a long-context phase extending to 512K tokens during pre-training. Supervised fine-tuning involves about 7.2 million samples, totaling roughly 100B tokens with ~65B trainable parameters. The training data includes 31.6% agentic and 68.4% non-agentic content, with software engineering comprising 69% of the agentic slice. Trajectories were generated using tools such as OpenHands, SWE-agent, Terminus-2, MiniSWE, Codex, and Goose. Quality control was conducted using GPT-OSS-120B and Gemma 4 as judges, along with SHA-256 deduplication for the tools and messages fields. Post-training involves a multi-stage, multi-environment reinforcement learning process, with each stage being an asynchronous GRPO run that warm-starts from the previous checkpoint. The order of training includes RLVR, skill boosters, SWE, Terminal, Search, and RLHF. The agentic RL block is exclusive to the 8B and 30B models, while the 3B model focuses on foundational RL and alignment. Training was conducted using NeMo-RL and NeMo-Gym on an NVIDIA GB200 NVL72 cluster hosted by CoreWeave. Additional support comes from 1 trillion synthetic code tokens generated by IBM’s CodeAlchemy pipeline and a speculative decoding layer for faster inference. IBM’s Granite 4.2 models are available for download, with detailed technical documentation provided on the IBM Research blog, a technical write-up, and a GitHub repository.