Granite Switch 4.1 builds on IBM's modular approach to model development. The new 8B and 30B models ship with a general-purpose decision adapter, a specialized safety-focused adapter dubbed Guardian, and Granite-Speech-5.0-Turbo-CTC, the fastest speech recognition model on the OpenASR leaderboard. All of these run together from a single endpoint while using roughly the same GPU memory as the base model alone.
The design uses cache-efficient Activated LoRAs, so the KV cache from one generation can be reused across adapters. That keeps switching costs low and enables tight workflows: the 30B variant completed a five-step RAG workflow faster than Jev-27B, and a custom 3B model made split-second Doom gameplay decisions at 60–100 decisions per second, compared with Jev's reported 10. The 3B model can also chat and narrate while playing.
Users are not limited to the bundled configuration. The project's Granite Libraries provide a repository of adapters that anyone can mix and match, and IBM plans to release an Adapter SDK for training custom adapters. The authors also note that specialized adapters often beat generalist decision models on specific classification tasks, so combining a broad decision adapter with targeted specialists can provide both coverage and accuracy at little extra cost.