Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

Granite Switch 4.1 introduces a modular AI model that integrates chat, decision-making, and speech-to-text functions into a single endpoint. The new models are configured with a general Decision aLoRA, a specialized Guardian aLoRA for safety tasks, and an embedded automatic speech recognition (ASR) model. This architecture utilizes cache-efficient Activated LoRAs (aLoRA) to reuse the base model's KV cache, enabling rapid switching between tasks while minimizing VRAM usage. The framework is extensible, allowing users to compose their own models by combining various adapters from the Granite Libraries to suit specific needs.