IBM releases Granite 4.2 reasoning models at 3B, 8B and 30B under Apache 2.0
IBM's Granite 4.2 is a dense, decoder-only reasoning family at 3B, 8B and 30B, pre-trained on roughly 15 trillion tokens with a 512K context, with the 30B reporting 57.00 on SWE-bench Verified and 89.17 on AIME25. The 8B and 30B also went through agentic RL in sandboxed software-engineering, terminal and web-search environments, putting an Apache 2.0 model with a thinking switch and OpenAI-compatible tool calling into existing agent harnesses.
Source: huggingface.co ↗
Every 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.
Granite Team, IBM
Why this matters
- → Apache 2.0 open reasoning models with tool use reduce reliance on proprietary APIs
- → 512K context and agentic RL in real environments bridge gap to frontier capability
- → Decoder-only architecture across 3B–30B sizes enables wider deployment than comparable models
Open reasoning with teeth