Granite Switch 4.1 adds decision adapters and speech-to-text to 8B and 30B models on one endpoint

Granite Switch 4.1 ships 8B and 30B models that combine a base Granite model, a general Decision aLoRA, a Guardian safety adapter, and Granite-Speech-5.0-Turbo-CTC for transcription, reportedly using about the same GPU VRAM as the base model alone. The authors say Activated LoRA adapters can reuse the base model KV cache, enabling chat, decision, and speech-to-text requests from one hosted endpoint; a demo transcribes audio, summarizes it, screens harmful content, and checks 15 user-defined flags in real time.

For builders of speech intake, moderation, or RAG workflows, the concrete serving advantage is fewer separate model endpoints because adapters and ASR are co-located and can share cache. The trade-off is coupling to the Granite Switch stack—vLLM needs a plug-in, and the article’s performance comparisons are vendor-reported, with more details and data still promised for upcoming weeks.