Granite 4.2 LLMs: How They're Built
Hugging Face released Granite 4.2, a reasoning-focused LLM family with 3B, 8B, and 30B models, featuring extended 512K-token context windows and multi-stage reinforcement learning for agentic tool use.
Hugging Face introduced Granite 4.2, its first dense decoder-only reasoning model family, available in three sizes: 3 billion, 8 billion, and 30 billion parameters. Each model was pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy that extends the context window to 512,000 tokens. The training pipeline includes supervised fine-tuning on chain-of-thought, reasoning, and agentic-trajectory data, followed by a multi-stage reinforcement learning process. All models support a thinking or non-thinking mode, with a low-effort intermediate mode for simpler queries, and are released under the Apache 2.0 license.
The 8B and 30B models undergo additional agentic reinforcement learning, enabling them to operate as agents by calling tools, editing and running code, using terminals, and performing web searches within sandboxed environments. All three models support native tool calling via an OpenAI-compatible endpoint, facilitating integration with agentic frameworks such as OpenCode and OpenHands without additional glue code. The models are also compatible with SGLang, with deployment recipes available in the SGLang cookbook.
Granite 4.2’s pre-training spans five phases, shifting from broad web-scale data to curated high-quality sources while progressively extending the context window. Supervised fine-tuning combines agentic and non-agentic data totaling about 7.2 million samples, or roughly 100 billion tokens, with rigorous quality control measures including LLM-based judging and deduplication. The agentic data covers domains such as software engineering, tool calling, and terminal use, while the non-agentic data includes instruction following, coding, and reasoning tasks.
The reinforcement learning pipeline consists of multiple stages targeting specific capabilities, including math, code, science, and tool use, followed by software engineering, terminal operations, and web search for the larger models. Each stage uses asynchronous Group Relative Policy Optimization (GRPO) with truncated importance sampling to handle policy drift, and concludes with reinforcement learning from human feedback (RLHF). The 30B model undergoes an additional agentic coding-focused fine-tuning phase to enhance its software engineering capabilities.