IBM ships Granite Switch 4.1, bundling chat, decision and speech models in one endpoint

IBM published Granite Switch 4.1 on October 15, two models (8B and 30B) that fold several capabilities into a single hosted endpoint.

What is inside

Each model combines the base language model with a new Decision adapter for general-purpose decisions, a Guardian adapter for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC, which IBM calls the fastest speech recognition model on the OpenASR Leaderboard. Together they take roughly the same GPU memory as the base model alone.

The adapters are Activated LoRAs, which let the KV cache from a base-model generation be reused for later adapter-based decisions. IBM says that, combined with recent vLLM contributions, lets one endpoint switch quickly between chat, decision and spoken requests. It also claims the Guardian adapter outperforms several named decision models on safety tasks; this is IBM’s own comparison.

Why it matters

Consolidating capabilities into one deployment reduces the number of GPUs enterprises need for chat, moderation and transcription workflows.