IBM's Granite Switch 4.1 Unifies Chat, Decisions, and Speech in a Single Endpoint

On October 16, 2026, IBM released the Granite Switch 4.1 model family, available in 8B and 30B variants. The models integrate a general-purpose decision aLoRA, a specialized Guardian safety decision aLoRA, and Granite-Speech-5.0-Turbo-CTC, the fastest speech recognition model on the OpenASR leaderboard, enabling seamless switching among conversational, decision-making, and speech transcription tasks from a single hosted endpoint.

IBM said the Granite Switch 4.1 models consume roughly the same GPU VRAM as the base model alone, delivering more capabilities with fewer GPUs. Because the Switch adapters are built as cache-efficient activated LoRAs, the KV cache from a base model generation can be reused for subsequent adapter-based decisions. Combined with recent vLLM contributions to speed up low-rank adapter inference and a co-located ASR encoder, the models can move quickly between chat, decision, and spoken requests.

Granite-Switch-4.1-8B can transcribe audio, generate summaries, screen for harmful content, and test 15+ generic flags in real time on a single endpoint. The Granite Switch project, introduced earlier this year, takes a modular approach to model development. IBM plans to release an Adapter SDK for training adapters in Mellea and encourages the community to compose custom Switch models using adapters from Granite Libraries.