Granite Switch 4.1 Overview
Released on October 15, 2026, Granite Switch 4.1 enhances a modular model approach, allowing a single endpoint to efficiently manage chat, decision-making, and speech-to-text tasks.
Key Components of Granite Switch 4.1
The Granite-Switch-4.1-8B and Granite-Switch-4.1-30B models are equipped with three main components:
- Decision aLoRA: Designed for general-purpose decision-making, exemplified by v1/systemone.
- Guardian aLoRA: A specialized adapter that reportedly outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash in safety-related decision tasks.
- Granite-Speech-5.0-Turbo-CTC: Claimed to be the fastest ASR model on the OpenASR Leaderboard, facilitating transcription tasks.
These components are integrated with the base model, ensuring that the combined model occupies a similar amount of GPU VRAM as the base model alone.
Cache Efficiency and Performance
The Switch adapters are designed as cache-efficient Activated LoRAs, allowing the KV cache generated from a base model to be reused for subsequent adapter-based inferences. This design, along with recent vLLM contributions, enhances the speed of low-rank adapter inferencing.
Real-Time Capabilities
Granite Switch 4.1 can seamlessly switch between chat, decision, and spoken requests at high speeds. In a demonstration, the granite-switch-4.1-8b model transcribes audio, generates summaries, screens for harmful content, and evaluates 15 user-defined flags—all in real time on a single endpoint.
Granite Libraries and Customization
Granite Libraries continues to provide specialized decision adapters that can be integrated into custom Granite Switch models. A performance comparison indicates that Granite-Switch-4.1-30B, utilizing a frozen base Granite model and four specialized aLoRAs, completed a five-step RAG workflow faster than Jev-27B, which uses a frozen Qwen base and one generic decision LoRA.
Custom Model Development
The team has developed a custom 3B parameter Granite Switch model capable of making rapid decisions while navigating the game Doom, achieving 60–100 decisions per second, compared to Jev’s reported 10 decisions per second. This model also provides narration and responses during gameplay.
Future Developments
The team plans to release an Adapter SDK for training adapters in Mellea, along with additional tutorials and support materials for users interested in customizing their Granite Switch models.
Original source: Hugging Face Blog