IBM launches Granite 4.2 reasoning models in 3B, 8B and 30B sizes

IBM released Granite 4.2 on August 25, 2026, a family of three open-weight language models that reason step by step before they answer. The sizes are 3B, 8B and 30B parameters, and every one of them ships under an Apache 2.0 licence. IBM Research says the family is purpose-built for the agentic workflows that enterprise use cases require, because reasoning helps the models plan before they act.

The capability split sits in post-training rather than in the architecture. As Ars Technica reported, the 8B and 30B variants go through an agentic reinforcement learning block that the 3B skips, training them on terminal use, web search and external tools. The 3B calls tools too, just without that specialised stage. That matters because all three look interchangeable on a spec sheet and behave differently inside an agent loop.

IBM’s own evaluation table shows where that line falls. The 8B model card reports no score at all for the 3B on the agentic coding benchmarks, while the reasoning benchmarks are scored across the board.

Benchmark	3B	8B	30B
SWE Bench Verified	NA	47.67	57
Terminal-Bench 2.1	NA	20.56	29.24
AIME25	78.33	86.67	89.17
GPQA	54.80	64.14	66.41

Granite 4.2 is the reasoning-focused release of the Granite language-model family.

Granite Team, IBM, via Hugging Face

Even so, reasoning costs something, and Ars Technica is blunt about the price. Chain-of-thought means more rigorous answers in some cases, but often slower response times and higher compute demands. IBM’s answer is a set of switches: the model card documents a full thinking mode, a non-thinking mode, and a low-effort mode that spends a short reasoning budget on easy questions. Whether you pay that cost locally or by the token is the same decision covered in our comparison of a hosted API against a self-hosted model.

The training run is documented in more detail than most releases bother with. IBM’s technical blog on Hugging Face describes a multi-stage reinforcement learning pipeline built on asynchronous Group Relative Policy Optimization, with separate stages for math, code, science, tool use, software engineering, terminal use and web search. IBM Research adds that the models were trained on 1 trillion tokens of synthetic code from its CodeAlchemy pipeline. A speculative decoding layer lets them output text faster while serving more users, which IBM frames as an operating cost argument rather than a benchmark one.

Two speech models landed the same day. Granite Speech 5.0 Turbo CTC and a non-commercial version come in at 470 million parameters, and IBM says they carry no LLM backbone, which is what keeps them small enough for laptops and phones. IBM puts the current speed leaders on the Hugging Face Open ASR leaderboard at roughly 6,000 RTFx, and says its own testing found Turbo CTC closer to 12,600 on a single H200 GPU. But that second figure is IBM’s internal measurement and not a leaderboard result, so treat it as a claim until someone reproduces it.

The Granite 4.2 repository lists quantized variants for each size alongside the full checkpoints, so the practical question is what these do on hardware you already own, which starts with picking a local runtime. IBM also says it’s working with Hirundo on machine unlearning to reduce undesirable outputs without retraining a model from the start, and no result is attached to that yet. The thing worth watching is whether the 30B’s agentic coding scores survive contact with an evaluation harness IBM didn’t build.