Granite 4.2 LLMs: How They're Built
Hugging Face ● Covered by 5 sources
IBM and Hugging Face laid out Granite 4.2, a new reasoning model family in 3B, 8B, and 30B sizes. It adds thinking modes, tool use, and a 512K context window, all under Apache 2.0.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Granite 4.2 is IBM’s latest swing at a reasoning model family, and this time the company is making the reasoning part explicit. The release comes in three dense, decoder-only sizes: 3B, 8B, and 30B. All three share the same architectural design and training pipeline, but the bigger two go further on agent behavior.
Each model is trained from scratch on roughly 15 trillion tokens. The pre-training recipe moves through five phases, starting broad and then tightening the data mix, with the last phase pushing the context window all the way to 512K tokens. After that comes supervised fine-tuning on a mixture of agentic and non-agentic data, followed by a multi-stage reinforcement learning pipeline.
That SFT mix is large: about 7.2 million samples, or roughly 100 billion tokens, with about 65 billion trainable. The agentic side leans heavily toward software engineering, but it also includes tool calling, terminal use, math, search, and action. The non-agentic side covers instruction following, coding, math, multilingual text, science, reasoning, and safety. IBM says the data is filtered, normalized into OpenAI Chat format, judged by GPT-OSS-120B and Gemma 4, then deduplicated with SHA-256 hashes.
The 8B and 30B models get the most interesting treatment after SFT. They go through an agentic RL block that teaches them to call tools, edit and run code, drive a terminal, and search the web in real sandboxed environments. All three models support native tool calling, and they can switch between thinking and non-thinking modes, with a low-effort thinking mode sitting in between for easier questions.
The RL side is staged rather than monolithic. IBM runs separate GRPO stages for verifiable rewards, skill boosters, software engineering, terminal use, search, and final RLHF, warm-starting each stage from the previous checkpoint. The 3B model takes a shorter path without the agentic block. Everything ships under Apache 2.0, which is the cleanest possible answer to the usual “closed or open?” debate: the weights may think, but the license doesn’t.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI that actually matters: less theater, more training recipe. Open models don’t win by yelling about vibes; they win by shipping the boring stuff well, and Granite 4.2 is plainly trying to do that. The real tell is the agentic RL on real sandboxed environments — that’s the sort of unglamorous grind closed labs love to keep to themselves.
Read more about this at: Hugging Face