IBM releases Granite 4.2 reasoning models
IBM has launched Granite 4.2, its first family of dense reasoning models featuring a switchable thinking mode and advanced reinforcement learning to act inside live software environments.
IBM released its Granite 4.2 family of dense, decoder-only reasoning language models on August 25, 2026. Published under the Apache 2.0 license, the models are available in 3B, 8B, and 30B parameter sizes. The training pipeline began with pre-training from scratch on approximately 15 trillion tokens, utilizing a five-phase schedule to stretch the context window to 512K tokens. This was followed by supervised fine-tuning on about 7.2 million samples of chain-of-thought, reasoning, and agentic-trajectory data.
The core of the release is a multi-stage reinforcement learning pipeline using asynchronous group-relative policy optimization (GRPO). All three sizes underwent foundational reinforcement learning on verifiable rewards, such as math problems and code graded by unit tests. However, only the 8B and 30B models advanced to the agentic block, where they acted inside live environments. This included editing repositories via the OpenHands harness, executing commands in a live shell terminal with up to 64 turns, and answering multi-hop search queries. These environments connect through NeMo-Gym, while NeMo-RL drives the training loop on Megatron-Core with vLLM generation. Training was conducted on an NVIDIA GB200 NVL72 cluster hosted by CoreWeave, featuring a 72-GPU NVLink domain and 400 Gb/s InfiniBand fabric.
IBM evaluated the models using a framework built on the NeMo Evaluator SDK. On the SWE-Bench Verified test, the 30B model scored 57.00 and the 8B scored 47.67. On Terminal-Bench 2.1, the 30B scored 29.24 and the 8B scored 20.56. For AIME25 math, the scores were 89.17 for the 30B, 86.67 for the 8B, and 78.33 for the 3B. On GPQA science, the 30B reached 66.41, the 8B hit 64.14, and the 3B scored 54.80. Finally, on the RULER long context benchmark at 128K, the 30B scored 81.38, the 8B scored 71.41, and the 3B scored 55.30.
Practitioners can control the models via three operating modes: a default thinking mode, a direct non-thinking mode, and a low-effort setting. Native tool calling is integrated directly into the chat template, allowing integration with vLLM or SGLang. The 30B model was post-trained from the Granite 4.1 30B base. The models support 12 languages, including English, German, Japanese, Arabic, Korean, and Chinese. IBM also released quantized versions, including FP8, NVFP4, and MXFP4 calibrated on 2,000 samples, alongside 14 GGUF formats via llama.cpp.
This is our own summary of reporting by Unite.AI
