IBM shipped a reasoning model in three sizes, and the smallest fits a laptop

IBM published Granite 4.2 on August 25, its first family of dense, decoder-only reasoning models. Three sizes — 3 billion, 8 billion and 30 billion parameters — each pre-trained from scratch on roughly 15 trillion tokens in a five-phase run that extends the context window to 512K, then post-trained through a multi-stage reinforcement learning pipeline. All of it ships under Apache 2.0.

What the models do differently

Every model in the family has a thinking and a non-thinking mode, plus a low-effort mode that spends a short reasoning budget on easy questions instead of the full chain. Tool calling is native rather than bolted on through prompting, and for the 8B and 30B models the reinforcement stage includes agentic RL: the model learns to act with tools inside sandboxed environments, not only to answer.

IBM documents the harnesses it expects the models to be driven from — OpenCode, Pi and OpenHands — which is a clearer signal of intent than any leaderboard position. This family is aimed at agents that run somewhere you control.

The formats are the news

Alongside the full-precision weights, IBM published FP8, NVFP4 and MXFP4 conversions made with LLM Compressor, and GGUF conversions made with llama.cpp itself. The GGUF set runs from Q8_0 down to Q2_K, which means the model is ready for the runners most people actually use — llama.cpp, and everything built on it — without waiting for the community to quantize it.

That matters more than it sounds. An open model that arrives as bare weights is a project; an open model that arrives with a permissive licence and the file formats already cut is something a team can put behind a product the same week.

• Sizes: 3B, 8B, 30B — dense, decoder-only, reasoning.
• Context: 512K tokens, reached through a five-phase pre-training run.
• Licence: Apache 2.0, on every size.
• Quantized builds at release: FP8, NVFP4, MXFP4, and GGUF from Q8_0 to Q2_K.

What is not in the announcement

The write-up is a technical walkthrough of how the family was built, down to the training hardware: an NVIDIA GB200 NVL72 cluster hosted by CoreWeave, with a 72-GPU NVLink domain and 400Gb/s InfiniBand between racks. What it is not is an independent evaluation. Every result in it is IBM measuring its own model.

So the useful question for anyone choosing a model this week is not where Granite lands on a chart, but whether a 30B model under Apache 2.0, with 4-bit builds ready, does the specific job you need at a cost you can predict. That is a test you run yourself, and it is cheap to run precisely because the licence and the formats are not in the way.