IBM and Hugging Face released Granite Switch 4.1, a modular AI model that combines natural language chat, decision-making, and speech-to-text capabilities in a single deployable endpoint. Available in 8B and 30B parameter sizes, the models employ Activated LoRAs—a parameter-efficient technique that allows specialized adapters to share a base model's key-value cache. The companies claim this approach delivers multiple specialized capabilities while consuming roughly the same GPU memory as a standard base model alone. The release includes a general-purpose Decision adapter and a specialized Guardian adapter for safety-related classifications that reportedly outperforms competitors like Jev 1.13, Microsoft-Decision-1, and GPT-Luna-Decisions on specialized classification benchmarks. The model's third primary capability is Granite-Speech-5.0-Turbo-CTC, an automatic speech recognition system that IBM and Hugging Face claim ranks as the fastest on the OpenASR leaderboard. By combining cache-efficient adapter composition with recent vLLM optimizations for low-rank adapter inference, Granite Switch 4.1 can rapidly alternate between chat requests, decision tasks, and speech transcription on a single endpoint. The teams published a public demo showing the model simultaneously transcribing audio, summarizing content, screening for harmful material, and evaluating 15 user-defined classification flags in real time. The release emphasizes modularity and extensibility. Beyond the pre-configured Granite Switch models, IBM and Hugging Face published Granite Libraries—a repository of specialized task adapters that developers can combine to build custom models. The teams demonstrated a 3B-parameter Granite configuration that played the classic game Doom at 60–100 decisions per second while narrating gameplay and answering questions—roughly six times faster than Jev's reported decision throughput.