IBM releases Granite Switch 4.1 modular AI models

IBM has released Granite Switch 4.1, featuring 8B and 30B parameter models that combine multiple specialized tasks into one hosted endpoint. The models utilize Activated LoRAs (aLoRAs) to switch between natural language, decision-making, and speech-to-text without significantly increasing VRAM usage. This modular approach allows for high-speed switching by reusing the base model's KV cache.

Technical architecture of Granite Switch 4.1

• Includes a new Decision aLoRA for general-purpose decision-making.
• Features a Guardian aLoRA for safety-related decision tasks.
• Integrates Granite-Speech-5.0-Turbo-CTC, an ASR model optimized for speed.
• Uses Activated LoRAs (aLoRAs) to allow the KV cache from the base model to be reused for subsequent adapter-based tasks.
• Designed to occupy approximately the same amount of VRAM as the base model alone.

Why it matters

The modular design allows users to compose custom models by mixing adapters from Granite Libraries, and IBM plans to release an Adapter SDK for training in Mellea.