Granite Switch 4.1 Unifies Chat, Decisions and Speech on One Endpoint

Summary

After introducing Granite Switch last spring, IBM's Granite team released Granite-Switch-4.1-8B and Granite-Switch-4.1-30B on October 15, 2026. Each combines its base model with a general Decision aLoRA, such as v1/systemone, a Guardian aLoRA that averages better safety-decision results than Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash, and embedded Granite-Speech-5.0-Turbo-CTC, the OpenASR Leaderboard's fastest ASR model. The combined system uses approximately the base model's VRAM. Cache-efficient Activated LoRAs reuse its KV cache for later decisions, while vLLM inference improvements and a co-located ASR encoder accelerate switching among chat, decisions and transcription.

The 8B model can transcribe, summarize, screen harmful content and test more than 15 generic flags through one real-time endpoint. Granite Libraries supplies specialist adapters that can outperform generalists on focused classification tasks and add almost no KV-cache cost. In a five-step RAG race, a frozen Granite-Switch-4.1-30B with four specialist aLoRAs finished much faster than Jev-27B, which used a frozen Qwen base and one generic decision LoRA, although timings were not disclosed. Users can compose alternatives from Granite Libraries, with an Adapter SDK for training custom and third-party adapters in Mellea, tutorials and further support planned. A custom 3B Granite Switch model played Doom at 60 to 100 decisions per second, versus Jev's reported 10, while narrating and answering questions. Model cards, a speech-to-decisions demo and GitHub repository are available, with more details and data promised in the coming weeks.