IBM Releases Granite Switch 4.1 30B and 8B Modular LLMs
IBM has released Granite Switch 4.1 30B and 8B, modular large language models that integrate natural language chat, specialized decision adapters, and embedded speech-to-text into a single hosted endpoint. These capabilities are activated on demand using control tokens within the chat template. This modular architecture simplifies deployment by combining multimodal capabilities, decision-making, and text generation into a single endpoint model. It enables developers to build complex, versatile AI applications without managing separate services for transcription and reasoning. The 30B model utilizes a 30-billion parameter base with a 128K context window, alongside aLoRA decision adapters and an embedded 470M-parameter English speech-to-text encoder. Both models are released under the Apache 2.0 license and support backends like HuggingFace Transformers and vLLM.
## BACKGROUND
Modular LLMs use techniques like Low-Rank Adaptation (LoRA) or Activated LoRA (aLoRA) to dynamically activate specific sub-components or adapters during execution without reloading the entire network. Connectionist Temporal Classification (CTC) is a loss function used in speech recognition models like TurboCTC to enable fast and accurate alignment-free audio transcription.