Large language models are evolving before our eyes. In a few short years, they’ve gone from powering reactive chatbots to anchoring autonomous agents that can reason through problems, plan, execute, and self-correct when needed.  

LLMs are increasingly expected to complete tasks, not just generate responses. This can-do attitude is baked into IBM’s latest Granite language and speech models. 

Today, IBM is releasing its updated Granite 4.2 languages models. Available in 3B, 8B, and 30B parameter sizes, Granite 4.2 is purpose-built for the agentic workflows that today’s enterprise use cases require. These language models include “thinking” capabilities, native step-by-step reasoning that helps them plan before they act, weigh trade-offs before deciding on a path, and catch mistakes before they can play out in real life.  

In enterprise workflows, tasks can be ambiguous and involve many steps to complete. An AI model must be able to follow complex instructions, retrieve the correct information, choose the right tools, act in the right sequence, and verify the result. Reasoning helps Granite 4.2 navigate this process more reliably. 

Granite 4.2’s strong tool-calling and reasoning capabilities allow it to evaluate which applications to use and in what order rather than executing blindly. Software engineering agents built on Granite 4.2 can navigate codebases, handle multi-step development tasks, and operate seamlessly in terminal environments. 

Granite 4.2 is built for deployment across cloud, on-premises, and edge environments. Its dense architecture supports broad compatibility, and its multiple sizes give teams flexibility. Smaller models can handle high-throughput agentic tasks efficiently. Larger models can be reserved for deeper reasoning and more complex coding workflows. 

Because Granite 4.2 is released under an Apache 2.0 license, organizations can download, fine-tune, and put it into production without licensing restrictions.    

A new training structure

Those agentic capabilities didn't emerge from model scale alone. The Granite team redesigned the training process to better prepare models for real-world tasks. Building on the Granite 4.0 foundation models, the IBM Research team introduced an expanded, multi-stage reinforcement learning (RL) regimen to deepen their capabilities.

Training begins with supervised fine-tuning and progresses through several RL phases. The first stage, known as "foundational RL," was applied across all Granite 4.2 models, strengthening capabilities in mathematics, science, coding, reasoning, and tool calling. This stage combines verifiable rewards with reward-model-based evaluation, helping models learn both accuracy and higher-level quality signals.

For the 8B and 30B models, training continues with a specialized "agentic RL" phase focused on enterprise-style tasks, including software engineering, terminal-based coding, and search-driven workflows. Combined with reinforcement learning from human feedback (RLHF) alignment, this approach produces models better equipped for complex, multi-step agentic work.

Two further innovations influenced Granite 4.2’s coding and reasoning improvements. The models were trained on 1 trillion tokens of synthetic code generated using IBM’s CodeAlchemy pipeline. They were also put through an intermediate training step called mid-training that’s been shown to unlock more reasoning power. Additionally, these models have a speculative decoding layer that allows them to output text faster while serving more users. This improves the AI inferencing experience for users and helps cut operating costs for enterprises. 

The team is also working with Hirundo to improve model robustness by reducing undesirable outputs post-training. Hirundo's machine unlearning technology is going to be used to target and reduce undesirable model outputs, without the need to fully retrain the model.      

Speech models designed for the edge

As part of today’s release, IBM is already introducing two new speech models. These models are a structural leap from the previous 4.1 models, which is why the team decided to name them Granite Speech 5.0 Turbo CTC and 5.0 Turbo CTC NC. These models come in at just 470 million parameters, making them among the smallest models in the Granite family and well suited for deployment on laptops, smartphones, and other edge devices. 

CTC, or connectionist temporal classification, helps speech models map audio to text efficiently and learn directly from raw audio and text. Using CTC, these models are very efficient at streaming audio, meaning they’re adept at quickly transcribing audio.  

Unlike previous Granite Speech models, they have no LLM backbone. This keeps the model size minuscule while remaining efficient at ASR (automatic speech recognition) tasks. The current speed leaders on the Hugging Face Open ASR leaderboard have a processing throughput (RTFx) score of around 6,000, while in testing, researchers found that Granite Speech 5.0 Turbo CTC came in closer to 12,600 — on a single H200 GPU.  

The result is a speech model optimized for Check out a demo of the new Granite Speech models in action on Hugging Face.high-throughput transcription workloads. Reduced sampling lets you process voice much faster, meaning it can transcribe three hours of voice recordings in a second. It’s a model lightweight enough for a user to run when transcribing a video chat in real time, or while coding a speech app, right from their laptop. Another key application is the high-volume transcriptions needed for analyzing call center data at scale. 

The team is also releasing a non-commercial (NC) version of the model trained on restricted-use data. 

Built for what comes next

This release extends the Granite family with a clear goal: helping enterprises build agents that can reason, act, and adapt during real-life workflows. Now that AI systems are being asked to carry out tasks in the real world, our expectations have risen. It’s no longer enough to answer clearly and concisely. An AI system must be able to plan, call applications, and execute complex tasks in a reliable and consistent way — while staying light enough to actually use without breaking the bank.  

These latest models were built with exactly those requirements in mind. Through advances in reasoning, coding, tool use, and speech processing, it moves enterprise AI closer to systems that can reliably participate in real-world work. 

You can download the new Granite models from any of your preferred model repositories, including Hugging Face, Ollama, and GitHub.