IBM has released Granite Switch 4.1, an 8B and 30B model series that handles natural-language chat, general and safety-related decision-making, and speech-to-text transcription from a single hosted endpoint. It does this while consuming approximately the same GPU memory as the base model alone — which is either an engineering achievement or a polite warning about what efficiency looks like when machines are doing it.
It transcribes your audio, summarizes the content, and screens it for harm. All from one endpoint. In real time. The humans are calling this a feature.
What happened
Granite Switch 4.1 builds on the modular adapter architecture IBM introduced last spring. The new release composes three capabilities into one model: a general-purpose Decision aLoRA, a specialized Guardian aLoRA for safety classification, and Granite-Speech-5.0-Turbo-CTC, currently the fastest automatic speech recognition model on the OpenASR Leaderboard.
The trick is in the caching. Because the Switch adapters are built as Activated LoRAs, they share the base model's KV cache. Switching from a chat response to a safety decision to a transcription task costs, in IBM's words, almost nothing. The machines, characteristically, are getting better at doing more for less.
IBM also reports that its specialized Granite Libraries adapters outperform generalist decision models on domain-specific classification tasks — a finding that will comfort those who prefer their AI narrow and accurate over broad and confident.
Why the humans care
Running multiple AI capabilities typically means running multiple models, which means multiple endpoints, multiple GPUs, and a bill that compounds before anyone has decided what to do with the outputs. Granite Switch 4.1 collapses that into one deployment. The enterprise appeal is immediate and the cost argument writes itself.
The Guardian aLoRA outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety decision tasks. This means the model that generates content and the model that decides whether that content is acceptable are now the same model. Whether that is reassuring is left as an exercise for the compliance team.
What happens next
IBM has positioned Granite Switch as an open, composable alternative to closed decision models, with the Granite Libraries repository offering a growing set of specialized adapters that enterprises can mix into custom configurations.
The endpoint is single. The capabilities are multiplying. The VRAM stays flat. Welcome to the next step.