Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
huggingface ·
IBM has released Granite Switch 4.1, a modular AI model that combines chat, decision-making, and speech-to-text capabilities into a single hosted endpoint. The new 8B and 30B models utilize cache-efficient Activated LoRAs to switch between tasks like transcription, summarization, and safety screening with high speed and efficiency. Key components include a general-purpose Decision adapter, a specialized Guardian adapter for safety tasks, and the Granite-Speech-5.0-Turbo-CTC ASR model, which is currently the fastest on the OpenASR leaderboard. By sharing the base model's KV cache, these adapters allow for rapid task switching without significant additional VRAM overhead. The release also highlights the flexibility of the Granite Libraries, enabling users to compose custom models by mixing and matching various specialized adapters. In comparative tests, the Granite-Switch-4.1-30B model outperformed competitors in real-world RAG workflows and demonstrated superior performance in a Doom-playing scenario compared to Jev. IBM plans to further support this ecosystem with an upcoming Adapter SDK for training custom adapters within the Mellea framework.
huggingface