IBM has released Granite 4.2, a new generation of open language models built to power AI agents that can reason, use tools, write code, follow complex instructions, and process speech. The models are available in 3B, 8B, and 30B parameter sizes and are released under an Apache 2.0 license, which lets organizations download, fine-tune, and deploy them without licensing restrictions.
Native Reasoning for Enterprise Agents
Granite 4.2 introduces native step-by-step reasoning, described by IBM as “thinking” capabilities. The models plan before acting, weigh trade-offs, and self-check before executing. According to IBM, reasoning helps the models handle enterprise tasks that are often ambiguous and require multiple steps, including retrieving the right information, choosing tools, sequencing actions, and verifying results.
Training and Tool-Calling Strength
The strongest tool-calling and reasoning capabilities are concentrated in the 8B and 30B variants, which were trained on enterprise-style tasks such as software engineering, terminal-based coding, and search-driven workflows. All three sizes received a foundational reinforcement learning stage covering mathematics, science, coding, reasoning, and tool calling. The 8B and 30B models were then refined through a specialized agentic reinforcement learning phase, combined with reinforcement learning from human feedback (RLHF) for alignment.
Two further techniques contributed to the coding and reasoning gains. IBM trained the models on 1 trillion tokens of synthetic code generated through its CodeAlchemy pipeline and used an intermediate “mid-training” step that IBM says unlocks additional reasoning power. A speculative decoding layer speeds up inference, allowing the models to serve more users at lower operating cost.
Deployment Architecture
The models use a dense architecture and are designed to run across cloud, on-premises, and edge environments. Smaller variants can handle high-throughput agentic tasks, while larger models are positioned for deeper reasoning and more complex coding work.
Granite Speech 5.0 Turbo Models
IBM is also releasing two new speech models alongside Granite 4.2: Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC. At 470 million parameters, both are among the smallest models in the Granite family and are intended for laptops, smartphones, and other edge devices. They use connectionist temporal classification (CTC) to map audio to text without an LLM backbone, making them compact and efficient for automatic speech recognition (ASR).
IBM reports that the new speech models transcribe significantly faster than current leaders on the Hugging Face Open ASR leaderboard, which post throughput (RTFx) scores around 6,000. In IBM’s testing on a single H200 GPU, Granite Speech 5.0 Turbo CTC reached roughly 12,600. Reduced sampling enables the model to process large volumes of audio, including transcribing three hours of voice recordings in a second. IBM highlights use cases such as real-time transcription during video chats, building speech applications on a laptop, and high-volume transcription for call center analytics. A non-commercial (NC) variant trained on restricted-use data is also available.
Machine Unlearning Partnership
IBM is additionally partnering with Hirundo to apply machine unlearning technology, which targets and reduces undesirable model outputs without requiring full retraining.
Availability
The Granite 4.2 models are available on Hugging Face, Ollama, and GitHub. A demo of the new Granite Speech models is available on Hugging Face.