3.66B, 8.79B, 29.28B dense (marketed as 3B, 8B, 30B)
131,072 tokens in released configs; IBM describes long-context extension to 512K
Apache License 2.0
Official open-weight reasoning-focused Granite release spanning three dense decoder-only models marketed as 3B, 8B, and 30B. IBM's Hugging Face collection reports exact parameter counts of approximately 3.66B, 8.79B, and 29.28B, while the released GraniteForCausalLM configurations use grouped-query attention, SwiGLU feed-forward layers, rotary position embeddings, and 131,072 maximum position embeddings. The official repository describes Granite 4.2 as post-trained on Granite 4.1 base checkpoints, while IBM's technical blog describes an underlying five-phase pretraining program of approximately 15 trillion tokens culminating in 512K long-context training; the published model cards consequently characterize 128K as native context with extension to 512K. All three models provide explicit chain-of-thought reasoning with full thinking, non-thinking, and low-effort modes, plus native OpenAI-compatible tool calling. Previous assistant reasoning is stripped from multi-turn history by default to conserve context, but the chat template exposes truncate_history_thinking=False for applications that want to preserve complete reasoning history. Supervised fine-tuning uses approximately 7.2 million samples and 100 billion tokens spanning reasoning, code, math, multilingual, safety, and agentic trajectories, followed by multi-environment GRPO and RLHF. The 8B and 30B models receive additional agentic reinforcement learning across sandboxed software-engineering, terminal, tool-use, and web-search environments. IBM reports 89.17 on AIME25, 75.77 on LiveCodeBench v6, 57.0 on SWE-bench Verified, 33.29 on SWE-bench Pro, 29.24 on Terminal-Bench 2.1, and 81.38 on RULER 128K for the 30B model. Official deployment artifacts include BF16 weights and IBM-produced GGUF, FP8, NVFP4, and MXFP4 quantizations, with serving guidance for Transformers, vLLM, SGLang, OpenCode, Pi, and OpenHands.