Hugging Face releases Granite Switch 4.1 with modular decision and speech adapters

Hugging Face released Granite Switch 4.1, a set of modular models that combine chat, decision-making, and speech-to-text capabilities into a single hosted endpoint. The 8B and 30B variants—Granite-Switch-4.1-8B and Granite-Switch-4.1-30B—include a new Decision adapter and Granite-Speech-5.0-Turbo-CTC, the latter being the fastest automatic speech recognition model on the OpenASR Leaderboard.

The Decision adapter is implemented as an Activated LoRA (aLoRA) for general-purpose decision-making, while a specialized Guardian aLoRA targets safety-related decision tasks. According to the announcement, the Guardian aLoRA outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on average in these evaluations.

Granite-Speech-5.0-Turbo-CTC is embedded for transcription and is positioned as the fastest ASR model on the OpenASR Leaderboard. The models are designed to share the base model’s VRAM footprint and reuse the KV cache generated from a base model generation for subsequent adapter-based decisions, enabling more capabilities with fewer GPUs.

Hugging Face reports that the Switch adapters were built as cache-efficient Activated LoRAs and that recent vLLM contributions to speed up low-rank adapter inference, combined with a co-located ASR encoder, allow the models to switch rapidly between chat, decision, and speech requests. A live demo space, Speech-to-Decisions, demonstrates real-time transcription, summarization, harmful content screening, and 15 other user-defined flags running on a single endpoint.

The release also emphasizes composability: users can build custom Granite Switch models by mixing adapters from Granite Libraries, which include specialized decision and task adapters. Hugging Face notes that adding specialists alongside a generalist decision adapter incurs minimal overhead because adapters share the base model’s KV cache, providing broad coverage with higher accuracy on targeted tasks.

In an internal comparison, Hugging Face reports that Granite-Switch-4.1-30B (frozen base Granite plus four specialized aLoRAs) completed a real-world five-step RAG workflow faster than Jev-27B (frozen base Qwen model plus one generic decision LoRA).

Hugging Face also announced plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and Granite Switch support, enabling users to extend models with custom or third-party adapters for specialized classification or other tasks. The company suggests such adapters could even be used to teach a 3B parameter Granite Switch model to play Doom, citing performance of 60–100 decisions per second compared to Jev’s reported 10 decisions per second, and the ability to narrate and respond to questions during gameplay.