IBM puts chat, safety checks and speech transcription in Granite Switch 4.1
Granite Switch 4.1 combines a language model, decision adapters and speech recognition. IBM says its 8B and 30B packages need approximately their base models' GPU memory and reuse conversation context.
In this article
One deployment for several kinds of request
IBM's October 15 announcement packages capabilities that an application might otherwise serve separately. A deployed Switch model can answer a conversational request, evaluate a decision and transcribe speech through the same endpoint. The memory-efficiency claim comes from IBM; it is not an independent measurement of every deployment.
The 8B model card explains what a decision means here: the application supplies a situation, a question and labeled alternatives, and the Decision adapter assigns probabilities to those choices. An adapter is a small set of additional model weights for a task. The accompanying Guardian adapter checks risks such as harmful content and attempts to bypass safety controls.
Reusing work when the task changes
The project explains that activated LoRA, or aLoRA, adapters begin applying their specialized weights at a designated control token. They can reuse the base model's key-value cache: stored intermediate calculations for text already processed. This avoids rebuilding that earlier context each time a compatible specialist is invoked. Developers can also compose their own compatible adapter selection into a deployable model rather than relying only on IBM's predefined bundle.
IBM reports a 50.1 overall score for the 30B configuration on the public, 1,068-row fast subset of Decision Index v0.3, on its 0–100 scale. This used a fixed system prompt, a situation processed by the frozen base model and a decision question whose options were scored from the first output token. It is a defined benchmark result, not a general percentage of real-world decisions the system gets right.
Speech first becomes text
The audio documentation labels this path alpha. A speech recognizer transcribes the recording, then passes ordinary text to the language model. The default recognizer handles English and produces lowercase text without punctuation. Vocal tone and emotion are lost, transcription errors carry forward, and long transcripts consume the model's context window. Hosting both components together does not turn them into a jointly trained speech-understanding model.
The same model card also says this bundle omits the library's factuality-checking and policy-guardrail adapters. Guardian does not replace application-specific safety testing, and decision probabilities should inform a choice rather than be treated as certified classifications in unfamiliar domains.
