IBM Releases Granite 4.2 Models for Reasoning, Coding and Agentic Workflows

What IBM released

On August 25, 2026, IBM released Granite 4.2 language models in 3B, 8B and 30B parameter sizes for enterprise agentic workflows. IBM says the models combine native step-by-step reasoning with tool calling, coding, instruction following and multi-stage task execution. The Granite 4.2 language models are released under the Apache 2.0 license.

Agent-focused training

IBM redesigned the training process around a multi-stage reinforcement learning regimen. All Granite 4.2 models receive a foundational RL stage aimed at mathematics, science, coding, reasoning and tool use. The 8B and 30B variants add a specialized agentic RL stage for enterprise-style tasks such as software engineering, terminal-based coding and search-driven workflows, followed by RLHF alignment.

Coding and inference changes

IBM says the models were trained on one trillion tokens of synthetic code produced by its CodeAlchemy pipeline and use an intermediate mid-training stage intended to strengthen reasoning. Granite 4.2 also includes speculative decoding to increase generation speed and serving efficiency.

New Granite Speech models

The same release introduces Granite Speech 5.0 Turbo CTC and a non-commercial variant. These 470-million-parameter speech models remove the LLM backbone used in previous Granite Speech systems and focus on efficient streaming automatic speech recognition. IBM reports an internal throughput result around 12,600 RTFx on one H200 GPU and says the model can support high-volume transcription and edge-oriented deployment.

Availability

The Granite 4.2 models are released and downloadable through multiple model and inference platforms, including Hugging Face, GitHub, Ollama and IBM watsonx. IBM's benchmark and throughput figures should be read as vendor-reported results until independently reproduced under comparable conditions.