IBM has released Granite Switch 4.1.1, an updated version of its modular large language model (LLM) designed for enterprise applications. This latest iteration enhances multilingual speech-to-text capabilities and refines its decision-making adapter, building upon the modular ‘Switch’ architecture that allows a single endpoint to handle diverse AI tasks.
The Granite Switch 4.1.1 model integrates a base LLM with specialized adapters and an encoder, enabling it to seamlessly transition between natural language responses, general and specialized decisions, and speech-to-text processing. These adapters are activated on demand via control tokens within the chat template, offering flexibility for developers.
According to details received by The Chenab Times, the update introduces significant improvements to its speech-to-text functionality. The embedded Granite Speech 5.0 TurboCTC encoder now supports transcription in English, French, German, Spanish, Portuguese, Italian, and Japanese. Furthermore, the transcripts are produced with native punctuation and capitalization, reducing the need for post-processing before they are fed into downstream models.
Enhanced Decision-Making and Modularity
Revision 4.1.1 also features updated weights for the decision adapter, which is built using aLoRA technology. This adapter has been evaluated on the Jev Decision Index v0.3 dataset, showing improved overall performance. The modular design of Granite Switch is a key feature, allowing for efficient updates and task-specific configurations. IBM plans to further empower developers by providing tutorials and an Adapter Software Development Kit (SDK), enabling them to create and iterate on their own adapters.
The Granite 4.1 family, first introduced on April 29, 2026, encompasses a range of models including language, vision, speech, embeddings, and Guardian safety models. These models are released under the Apache 2.0 license, making them available for both research and commercial use. The Granite 4.1 8B instruct model, for instance, has demonstrated capabilities comparable to or exceeding the previous Granite 4.0 32B Mixture-of-Experts model, while utilizing a simpler, more efficient dense architecture.
Granite 4.1 models are trained on approximately 15 trillion tokens through a multi-stage pre-training pipeline, with context lengths extending up to 512,000 tokens. This extensive training allows the models to process lengthy documents without performance degradation on shorter tasks. Key capabilities across the Granite 4.1 family include robust tool calling, improved instruction following, advanced coding, and mathematical reasoning.
The Chenab Times News Desk