Granite Switch, introduced last spring as a modular approach to model development, is now at version 4.1. The new Granite-Switch-4.1-8B and -30B models bundle the underlying Granite base model with a set of adapters so a single hosted endpoint can move between natural-language responses, general and specialized decision-making, and speech-to-text tasks. The included Decision aLoRA targets general-purpose decision-making (the post gives v1/systemone as an example), while a specialized Guardian aLoRA is reported to outperform Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks on average. For transcription, the models embed Granite-Speech-5.0-Turbo-CTC, described as the fastest ASR model on the OpenASR Leaderboard.

All three technologies are composed into one model that takes up approximately the same amount of GPU VRAM as the base model by itself, which the team says delivers more capabilities with fewer GPUs. The Switch adapters are cache-efficient Activated LoRAs: the KV cache created from a base model generation can be reused for any subsequent adapter-based decisions. Combined with recent vLLM contributions that speed up low-rank adapter inferencing and a co-located ASR encoder, the models can switch between chat, decision, and spoken requests with blazing fast speeds. As a concrete example, Granite-Switch-4.1-8B can transcribe audio, generate summaries, screen for harmful content, and test more than 15 generic flags, all running in real time on a single endpoint.

Decision models are not new to Granite. The previous Granite Libraries release already provided a repository of specialized decision-model adapters and other task adapters that can be composed into custom Granite Switch models. The team observes that while closed decision models such as Jev focus on generality, many specialized granitelib adapters outperform generalists on specialized classification tasks. Because adapters share the base model’s KV cache, adding specialized adapters alongside a generalist decision adapter costs almost nothing, giving both broad coverage and higher accuracy on the tasks that matter.

The post reports a head-to-head race between Granite-Switch-4.1-30B (frozen base Granite plus four specialized aLoRAs) and Jev-27B (frozen Qwen base plus one generic decision LoRA). Granite-Switch-4.1-30B completed a real-world, five-step RAG workflow much faster than its competitor.

Granite-Switch-4.1-8B and -30B ship batteries included, compiled with a predefined set of helpful adapters, but users can compose their own Granite Switch models by mixing and matching adapters from Granite Libraries. The project plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and Granite Switch support. Custom adapters could be specialized classification adapters for decision workflows or adapters for specialized language tasks. The authors even demo a custom 3B-parameter Granite Switch model that plays Doom: it makes split-second navigation decisions at 60-100 decisions per second, compared with Jev‘s reported 10 decisions per second, and unlike Jev it can narrate and answer questions while playing.

More details and data on Granite Switch are expected in the coming weeks. For getting started, the post links to the Granite Switch 4.1 model cards, a speech-to-decisions demo, the Granite Switch GitHub repository, the original Granite Switch and Granite Libraries release blog, and a Colab notebook for creating your own aLoRA-vs-LoRA race. The 4.1 release illustrates the project’s direction: rather than choosing between a generalist assistant, a specialist decision model, and an ASR system, a single composed model can offer all three with minimal extra VRAM overhead.