IBM Releases Granite 4.2 Under Apache 2.0 with Agentic Features

IBM releases its new generation Granite 4.2 under the Apache 2.0 license, scaled up to 30B parameters, along with the Granite Speech 5.0 Turbo CTC series. The company highlights agentic capabilities, an extended context window, and speed gains in speech recognition, while its internal benchmarks indicate an advantage for the 30B model on tooling tasks.

Apache 2.0 Distribution and Announced Performance

All Granite 4.2 and Granite Speech 5.0 Turbo CTC models are offered under the Apache 2.0 license and are available on Hugging Face, Ollama, GitHub, and other platforms. Granite Speech 5.0 Turbo CTC features 470 million parameters. IBM claims that these speech recognition models are twice as fast as the previous leaders on the Open ASR Leaderboard and can transcribe three hours of audio in one second. In internal benchmarks focusing on agentic tasks and tool usage, the Granite 4.2 30B model achieves the best results across all three tests.

Sizes, Training, and Context Management

Granite 4.2 is available in three sizes: 3 billion, 8 billion, and 30 billion parameters. IBM states that the training of these models was conducted from scratch on nearly 15 trillion tokens. They are capable of processing contexts of up to 512,000 tokens. The models feature "thinking" and "non-thinking" modes that allow resource usage to be tailored to the nature of the task, and a "low effort" mode is expected to be offered for simple queries.

Agentic Learning and Tool Compatibility

The Granite 4.2 8B and 30B variants benefit from "agentic RL" training, enabling them to learn how to use tools, write and execute code, and perform web searches in sandbox environments. All Granite 4.2 models offer compatibility with tool calls in OpenAI format and can be used with vLLM or SGLang inference servers.