IBM has released the Granite 4.2 family of language models in 3B, 8B, and 30B sizes under the Apache 2.0 license. Trained from scratch on approximately 15 trillion tokens, these models support context windows up to 512,000 tokens and can toggle between thinking and non-thinking modes to control compute per task. The 8B and 30B variants undergo agentic RL training, enabling them to use tools, write and run code, and search the web in real sandbox environments. Additionally, the new Granite Speech 5.0 Turbo CTC models contain just 470 million parameters and transcribe three hours of audio in one second. All models run on vLLM or SGLang and are available on Hugging Face, Ollama, and GitHub.
The inclusion of agentic capabilities directly addresses the growing need for models that can execute multi-step tasks without constant human supervision. By integrating tool use and code execution into the base training process, IBM aims to reduce the engineering overhead required to build autonomous agents. The Apache 2.0 licensing ensures broad commercial adoption without restrictive clauses, while the variable compute modes allow users to balance performance against resource usage.
* Models support OpenAI-format tool calling
* Speech models achieve twice the speed of previous leaders on the Open ASR Leaderboard
* Available immediately on Hugging Face and Ollama
