IBM open-sources Granite 4.2 reasoning models in 3B/8B/30B sizes under Apache 2.0

On August 25, IBM released Granite 4.2 open-source language models in 3B/8B/30B parameter sizes. Unlike earlier instruction-following Granite releases, Granite 4.2 is built around explicit reasoning: every model can emit a chain of thought before answering, and every model exposes a thinking/non-thinking switch plus a low-effort mode for simple queries.

All three models are decoder-only dense transformers, pre-trained from scratch on approximately 15 trillion tokens. The 3B uses 40 layers (embedding size 2560), the 8B uses 40 layers (4096), the 30B goes to 64 layers with an MLP hidden size of 32,768. The published architecture table lists a 131,072-token sequence length, while the five-phase pre-training run includes a phase extending to 512K tokens.

Granite 4.2 post-training uses a multi-stage, multi-environment RL chain. Each stage is a separate asynchronous GRPO that warm-starts from the previous checkpoint, with leave-one-out baseline instead of a value network, and truncated importance sampling to bound off-policy drift. Order is RLVR, skill boosters, SWE, Terminal, Search, then RLHF. The agentic RL block runs only on the 8B and 30B; the 3B takes foundational RL and alignment only.

All three models are released under Apache 2.0. Alongside the language models, IBM released two 470M-parameter Granite Speech 5.0 Turbo CTC models, twice as fast as previous leaders on the Open ASR Leaderboard, able to transcribe three hours of audio in one second. Models are available on Hugging Face, Ollama, GitHub, and other platforms.