IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models

IBM has released Granite 4.2 , a family of open reasoning language models in 3B, 8B, and 30B parameter sizes. Unlike earlier Granite releases, which were instruction-following assistants, Granite 4.2 is built around explicit reasoning. Every model can emit a chain of thought before answering, and every model exposes a thinking / non-thinking switch plus a low-effort mode that spends a short reasoning budget on easy questions. The models are decoder-only dense transformers, pre-trained from scratch on roughly 15 trillion tokens, then post-trained through a multi-stage reinforcement learning chain. For the 8B and 30B, that chain includes an agentic RL block where the model learns to edit code, drive a terminal, and run web searches inside real sandboxed environments. All three ship under Apache 2.0 . IBM also released two 470M-parameter Granite Speech 5.0 Turbo CTC models alongside the LLMs.

Yes, All three Granite 4.2 language models ship under Apache 2.0 , so download, fine-tuning, and commercial production use carry no licensing gate.

Granite 4.2 is a decoder-only dense transformer, not a hybrid or MoE design. Core components are Grouped Query Attention with 8 KV heads, RoPE with θ = 10,000,000, SwiGLU MLPs, RMSNorm (ε = 1e-5), untied input/output embeddings, and bfloat16 precision.

The 3B uses 40 layers at embedding size 2560. The 8B uses 40 layers at 4096. The 30B goes to 64 layers with an MLP hidden size of 32,768. The published architecture table lists a 131,072-token (128K) sequence length , while the five-phase pre-training run includes a long-context phase extending to 512K tokens. Pre-training covers roughly 15 trillion tokens from scratch.

Supervised fine-tuning uses about 7.2 million samples , roughly 100B tokens with ~65B trainable. The mixture is 31.6% agentic and 68.4% non-agentic , and software engineering is 69% of the agentic slice. Trajectories were generated across harnesses including OpenHands, SWE-agent, Terminus-2, MiniSWE, Codex, and Goose. Quality control used GPT-OSS-120B and Gemma 4 as judges, plus SHA-256 deduplication over the tools and messages fields.

Post-training is a multi-stage, multi-environment RL chain , not a single pass. Each stage is a separate asynchronous GRPO run that warm-starts from the previous checkpoint, with a leave-one-out baseline instead of a value network and truncated importance sampling to bound off-policy drift. The order is RLVR, then skill boosters, then SWE, Terminal, Search, then RLHF.

The agentic RL block runs only on the 8B and 30B . The 3B takes foundational RL and alignment only. That single design choice explains most of the capability gap across sizes. Training ran on NeMo-RL and NeMo-Gym over an NVIDIA GB200 NVL72 cluster hosted by CoreWeave.