Granite Switch 4.1 Unifies Speech, Decisions, and Chat

The Granite Switch project has launched its 4.1 update, introducing the Granite-Switch-4.1-8B and Granite-Switch-4.1-30B models. These architectures integrate a Decision aLoRA for general workflows like v1/systemone, a specialized Guardian aLoRA for safety tasks, and Granite-Speech-5.0-Turbo-CTC, which holds the top speed spot on the OpenASR leaderboard. By composing these tools into a single unit, the system delivers text responses, speech transcription, and classification from a single endpoint using roughly the same GPU VRAM as the base model alone.

The architecture relies on Activated LoRAs and speed enhancements contributed to vLLM to achieve low latency. Because adapters share the underlying KV cache, switching contexts incurs almost no performance penalty. In evaluations, the Guardian aLoRA outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety decisions. Furthermore, during a race over a 5-step RAG workflow, Granite-Switch-4.1-30B—built from a frozen Granite base and 4 specialized aLoRAs—finished faster than Jev-27B, which uses a frozen Qwen base and 1 generic decision LoRA. Practical implementations, such as the Speech-to-Decisions HF space, demonstrate Granite-Switch-4.1-8B handling live audio transcription, summaries, and 15 user-defined flags concurrently.

Engineers can customize these builds using Granite Libraries or prepare custom components with a future Adapter SDK in Mellea. Highlighting this responsiveness, a compact 3B parameter Granite Switch variant played the game Doom while generating 60 to 100 decisions per second, surpassing Jev's 10 decisions per second while answering questions and narrating gameplay in real time.