IBM has released Granite Switch 4.1 models that combine conversation, structured decisions and speech transcription behind one endpoint. Detailed in an October 15 announcement, the new packages use 8B and 30B Granite base models and are available under Apache 2.0.
The idea is to give one deployment several jobs. A developer could transcribe a spoken request, classify what it needs and generate a reply without maintaining a separate full language model for every step.
Three capabilities in one package
According to IBM’s announcement, both models include a general-purpose Decision adapter, a Guardian adapter for detecting safety-related risks, and an embedded speech transcription model.
The decision component scores a set of labelled options. That makes it useful for tasks such as routing a support request to the right team. Our earlier coverage of Cloudflare’s Clef decision models explains how this differs from asking a chatbot to write a free-form answer.
Granite Switch keeps ordinary chat available alongside those specialist functions. Applications select the adapter they need through the request’s chat-template settings.
Adapters share the base model’s work
An adapter is a smaller set of trained weights that changes how a base model handles a particular task. Granite Switch uses activated LoRA adapters, which turn on when needed.
The project’s GitHub documentation explains that these adapters share a compatible conversation cache. Earlier processing can be reused when the application switches tasks, reducing the need to process the same context again.
IBM says the combined package uses approximately the same GPU memory as its base model alone. That is IBM’s deployment claim, rather than an independent measurement of every workload.
The base model still sets much of the memory budget. As a rough calculation, eight billion parameters at BF16 precision take around 16GB for weights alone. Adapters, the speech component, working memory and conversation cache add to that. Our local AI memory guide covers those extra requirements.
Speech support has clear limits
The 8B model card lists a 128K context window and a 470-million-parameter speech encoder. The audio path first transcribes speech, then places the text into the language model’s conversation.
Transcription currently supports English only and produces lowercase text without punctuation. The speech encoder and language model have not been jointly trained, so a transcription mistake can affect the answer or decision that follows.
The published setup uses Hugging Face Transformers or vLLM with Granite Switch. Model cards list specific dependency versions, so users should follow the checkpoint’s setup instructions.
Downloads are available now. Developers can use IBM’s prepared packages or compose their own models with selected adapters through the project’s tooling.
