Granite Switch 4.1 bundles chat, moderation and speech into one model
IBM stacks decision-making and speech transcription onto a chat model as adapters that share its KV cache, so the extra capabilities cost almost no additional GPU memory.
IBM published Granite Switch 4.1 on 15 October: two Apache 2.0 models, 8B and 30B parameters, that handle chat, decision-making and speech transcription from one hosted endpoint instead of three separate models.
The base is an ordinary chat model. On top, IBM stacks adapters: a Decision adapter for general-purpose judgments, a Guardian adapter tuned for safety and content classification, and Granite Speech 5.0 Turbo CTC for transcription. Because the adapters share the base model’s KV cache, switching between capabilities adds almost no extra GPU memory. IBM says the Guardian adapter beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on its safety classification tests, and a 3B preview variant handles 60 to 100 decisions a second against Jev’s reported 10. Granite Speech 5.0 tops the OpenASR leaderboard for transcription speed. It deploys through a vLLM plugin or llama.cpp.
Why it matters: teams running separate models for chat, content moderation and transcription can collapse them onto one endpoint without extra GPU budget, and the weights are open under Apache 2.0 rather than locked behind someone else’s API.
The caveat: the benchmark comparisons are IBM’s own, and the smallest size, the 3B model, is still marked preview.