IBM Granite Switch 4.1 combines chat, decisions, safety, and speech on one endpoint

What is changing?

The IBM Granite team announced Granite Switch 4.1 in 8B and 30B sizes on October 15. The 8B model combines a Granite 4.1 base with a 128K-token context window, a general-purpose Decision aLoRA, a Guardian safety adapter, and a 470-million-parameter Granite Speech 5.0 TurboCTC encoder in one checkpoint. According to the model card, English speech recognition can be called through the same API without a separate speech server.

Activated LoRA adapters reuse the base model’s KV cache, while the speech encoder is co-located in the same deployment. IBM says the composition uses roughly as much GPU memory as the base model alone. In the official live demo, a single 8B endpoint transcribes speech, names the topic, screens for harmful content, and evaluates user-defined yes-or-no decisions in the same flow.

The weights are released under Apache 2.0. The 8B card documents Hugging Face Transformers and vLLM backends, while the launch article also describes llama.cpp support. The speech encoder and language model are not jointly trained, however; speech output is English-only and can be lowercase and unpunctuated. Production teams should independently test routing accuracy, latency, and safety decisions on their own workloads.