IBM has released Granite 4.2, a new family of dense, decoder-only reasoning language models distributed under an Apache 2.0 open-source license. Available in three parameter sizes (3B, 8B, and 30B), the series introduces native chain-of-thought deliberation, multi-tier reasoning controls, and sandboxed agentic reinforcement learning for open-weight deployments.
The release marks a architectural shift from IBM's earlier instruction-following Granite checkpoints toward explicit reasoning and autonomous tool interaction. All three variants share a common architecture and pre-training pipeline, while the larger 8B and 30B checkpoints incorporate specialized agentic post-training for environment execution.
Architectural Specifications and Pre-Training
The Granite 4.2 family uses a dense transformer architecture configured with Grouped Query Attention (GQA), Rotary Position Embeddings (RoPE) with a base frequency parameter of 10,000,000, RMSNorm normalization, SwiGLU activation functions, and untied input-output embeddings.
| Parameter | 3B Dense | 8B Dense | 30B Dense | | :--- | :--- | :--- | :--- | | Layers | 40 | 40 | 64 | | Embedding Dimension | 2560 | 4096 | 4096 | | Attention Heads (Query / KV) | 40 / 8 | 32 / 8 | 32 / 8 | | Head Dimension | 64 | 128 | 128 | | MLP Hidden Dimension | 8192 | 12800 | 32768 | | Base Sequence Length | 131,072 | 131,072 | 131,072 | | Maximum Extended Context | 512,000 | 512,000 | 512,000 |
Each model was trained from scratch across approximately 15 trillion tokens using a five-phase pre-training curriculum. Phases 1 and 2 established baseline language and code capabilities, phases 3 and 4 conducted data annealing on curated high-quality subsets, and phase 5 applied context extension scaling to reach 512,000 tokens.
Supervised Fine-Tuning and Agentic Trajectories
The supervised fine-tuning (SFT) corpus comprised roughly 7.2 million samples across 100 billion tokens (65 billion trainable tokens), divided into 31.6% agentic and 68.4% non-agentic data.
The agentic training subset focused on software engineering tasks (69.0%), function calling (12.1%), terminal operations (8.0%), mathematics (3.5%), search execution (0.8%), and computer actions (0.2%). IBM generated and formatted trajectories across multiple agent harnesses, including OpenHands, OpenCode, Terminus-2, SWE-agent, OpenResearcher, MiniSWE, OpenSeeker, EnvScaler, Gemini CLI, Hermes, Codex, and Goose.
Quality filtering utilized GPT-OSS-120B and Gemma 4 as automated judges to prune low-quality traces, hallucinated tool schemas, and ungrounded execution steps. For the 30B model, IBM added a second SFT phase specifically upsampling agentic coding data with a 16% replay buffer to prevent catastrophic forgetting.
Multi-Stage Reinforcement Learning and Serving
Following SFT, models underwent a chained multi-stage reinforcement learning pipeline. Successive RL runs targeted discrete competencies (mathematics, coding, instruction adherence, structured JSON output, software engineering, and terminal command navigation) by warm-starting from each preceding checkpoint.
For the 8B and 30B models, agentic RL trained models inside sandboxed code execution environments with real execution feedback. The architecture supports dynamic reasoning modes:
- Thinking Mode: Full chain-of-thought reasoning prior to final response generation.
- Non-Thinking Mode: Direct response output for low-latency instruction following.
- Low-Effort Mode: Constrained reasoning token budget for straightforward prompts.
Granite 4.2 emits function calls in standard OpenAI API formats and is supported out-of-the-box in vLLM, SGLang, and local terminal agent harnesses including Pi and OpenHands.
