Published · 9h agoBuild3 min read
Granite 4.2 ships a self-hostable reasoning tier under Apache 2.0, and its data says coding agent
IBM's 3B, 8B and 30B dense models all get a thinking switch and native tool calling, but only the two larger ones get agentic RL, and the tuning mixture leans hard on software engineering.
Written for builders.See today for builders
What happened
- IBM has released Granite 4.2 as three dense, decoder-only reasoning models at 3B, 8B and 30B, sharing one architecture and one training pipeline.
- All three carry a thinking / non-thinking switch, a low-effort thinking mode and native tool calling.
- Each was pre-trained from scratch on about 15 trillion tokens across five phases, the last extending the context window to 512K tokens.
- Only the 8B and 30B get the agentic RL stage, which trains tool use, code editing, terminal driving and web search inside real sandboxed environments.
- The supervised stage used roughly 7.2 million samples and 100 billion tokens, split 31.6 percent agentic and 68.4 percent non-agentic.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- capabilityTeams that cannot send prompts to a hosted endpoint now have a thinking-mode model in three sizes on one permissive license, runnable inside their own boundary.
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Granite 4.2 is IBM's first family of dense, decoder-only reasoning LLMs, released in three sizes: 3B, 8B and 30B, all sharing the same architectural design and training pipeline.
- [2]
All Granite 4.2 models are released under the Apache 2.0 license.
- [3]
Each model is pre-trained from scratch on approximately 15 trillion tokens using a five-phase strategy; phase 5 introduces long-context training, extending the context window to 512K tokens.
- [4]
Every Granite 4.2 model has a thinking / non-thinking switch, a low-effort thinking mode that spends a short reasoning budget on easy questions, and native tool calling.
- [5]
The 8B and 30B models additionally go through an agentic RL block that teaches them to call tools, edit and run code, drive a terminal and search the web inside real sandboxed environments.
- [6]
The SFT data mixture combines agentic (31.6%) and non-agentic (68.4%) data, totalling approximately 7.2 million samples, or roughly 100B tokens, of which about 65B are trainable.
Sources & coverage · 2 publishers
The reporting this story was synthesized from, earliest first. Every link goes to the original.
- huggingface.co10h agoGranite 4.2 LLMs: How They're Built
- thenewstack.ioFrederic Lardinois6h agoIBM’s new Granite 4.2 models add reasoning and stay dense
Additional citations
- Granite Team, IBM
