IBM has introduced the Granite 4.2 family of language models, available in three sizes: 3 billion, 8 billion, and 30 billion parameters. According to The Decoder, these models were trained on an extensive dataset of approximately 15 trillion tokens and support an impressive context window of up to 512,000 tokens, enabling them to process long-form content effectively.
The larger Granite models incorporate 'agentic reinforcement learning' techniques, allowing them to autonomously learn tool use and execute code, enhancing their versatility in practical applications. All versions of the Granite 4.2 models are released under the Apache 2.0 license, ensuring broad accessibility for developers and enterprises.
This release could be particularly relevant for Japanese markets, where advanced AI tools are increasingly integrated into financial services and trading platforms, potentially improving automation and decision-making processes in FX, crypto, and equities sectors.