IBM Granite combines chat, decisions and speech in Granite Switch 4.1

IBM Granite configured Granite-Switch-4.1-8B and Granite-Switch-4.1-30B with chat, decision-making and speech-to-text capabilities.[1] A single checkpoint supports specialized functions including retrieval-augmented generation (RAG), safety and explainability, activated on demand through chat-template control tokens.[2] The Decision aLoRA is designed for general-purpose decision-making.[1] IBM Granite says its Guardian aLoRA outperforms four named decision models on average on safety-related decision tasks.[1] The evidence does not provide the models’ names or the size of the reported margin.

IBM Granite says the combined technologies use approximately the same GPU VRAM as the base model alone, and that a base-model KV cache can be reused for later adapter-based decisions.[1] The company says its recent vLLM contributions speed up low-rank adapter inference.[1] Its Speech-to-Decisions demo uses one endpoint in real time to transcribe audio, summarize it, screen for harmful content and test 15 other user-defined flags.[1] IBM also says specialist adapters can be added alongside a generalist decision adapter at almost no cost because the adapters share the base model’s KV cache.[1]

IBM Granite says its 30B Switch model completed a real-world five-step RAG workflow much faster than a competitor.[1] It also reports 60–100 decisions per second for its model, compared with Jev’s reported 10 decisions per second.[1] These are company-reported comparisons; the evidence does not specify the competitor in the RAG test or provide test conditions. IBM says its Switch model can narrate and answer questions while playing, unlike Jev.[1] Users can compose their own Switch model by mixing and matching adapters from Granite Libraries.[1]

The 3B Preview model card says the model includes 12 adapters from the RAG, core and Guardian Granite libraries.[2] It lists a 128K-token context for the Granite 4.1 base model, an Apache 2.0 license, and support for Hugging Face Transformers and vLLM.[2] The model card gives May 5, 2026, as the Preview model’s release date.[2] IBM plans to release an Adapter SDK for training adapters in Mellea, with tutorials and Granite Switch support.[1] IBM says it released three libraries designed to support common enterprise workflows.[3]