---
title: "Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint"
description: "Granite Switch 4.1 models combine chat, general and specialized decisions, and speech-to-text from a single hosted endpoint. The 8B and 30B models include a new Decision aLoRA for general-purpose decision-making and a Guardian aLoRA that outperforms Jev 1.13 and other closed models on safety tasks. They also embed Granite-Speech-5.0-Turbo-CTC, the fastest ASR model on the OpenASR Leaderboard, enabling efficient transcription. All components are composed into one model using cache-efficient Activated LoRAs, taking up about the same GPU VRAM as the base model alone, and the KV cache from base generation can be reused for any subsequent adapter decisions. Granite Switch 4.1 can switch between chat, decision, and spoken requests quickly, and a 3B variant achieved 60-100 decisions per second playing Doom, compared to Jev’s 10, while also narrating and answering questions."
---

# Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

[Hugging Face](https://yomu.fyi/company/hugging-face) · Kate Soule, Subhajit Chaudhury, Hendrik Strobelt, Luis A Lastras · Oct 15, 2026

**Type:** Announcement

## Summary

Granite Switch 4.1 models combine chat, general and specialized decisions, and speech-to-text from a single hosted endpoint. The 8B and 30B models include a new Decision aLoRA for general-purpose decision-making and a Guardian aLoRA that outperforms Jev 1.13 and other closed models on safety tasks. They also embed Granite-Speech-5.0-Turbo-CTC, the fastest ASR model on the OpenASR Leaderboard, enabling efficient transcription. All components are composed into one model using cache-efficient Activated LoRAs, taking up about the same GPU VRAM as the base model alone, and the KV cache from base generation can be reused for any subsequent adapter decisions. Granite Switch 4.1 can switch between chat, decision, and spoken requests quickly, and a 3B variant achieved 60-100 decisions per second playing Doom, compared to Jev’s 10, while also narrating and answering questions.

## Context

IBM’s Granite Switch project, introduced last spring, enables modular model development by composing specialized adapters onto a base model, allowing a single endpoint to handle multiple task types.

## Approach / What changed

Granite Switch 4.1 uses cache-efficient Activated LoRAs (aLoRAs) for decision-making and safety, embeds the fast Granite-Speech-5.0-Turbo-CTC ASR encoder, and shares a base model’s KV cache across adapters, enabling chat, decision, and speech tasks from one endpoint with minimal VRAM overhead.

## Takeaways

- Granite Switch 4.1 combines chat, decision-making, and speech-to-text in one model using cache-efficient activated LoRAs.
- The Guardian aLoRA outperforms Jev 1.13, Microsoft-Decision-1, and GPT-Luna-Decisions on safety-related decision tasks.
- Specialized adapters from Granite Libraries can be added alongside generalist decision adapters with negligible cost due to KV cache sharing.

## FAQ

### What does Granite Switch 4.1 combine?

It combines chat (natural-language responses), decision-making (general and specialized), and speech-to-text (ASR) from a single endpoint.

### How does it achieve efficiency?

By using cache-efficient Activated LoRAs that share the base model's KV cache, all adapters take up roughly the same VRAM as the base model, and switching between tasks is fast.

### How does it compare to Jev?

Granite-Switch-4.1-30B completed a 5-step RAG workflow faster than Jev-27B, and a 3B variant achieved 60-100 decisions per second playing Doom compared to Jev's 10.

**Tags:** [ASR](https://yomu.fyi/topic/asr), [Granite](https://yomu.fyi/topic/granite), [LLMs](https://yomu.fyi/topic/llm), [LoRA](https://yomu.fyi/topic/lora), [RAG](https://yomu.fyi/topic/rag)

- Source: [Hugging Face](https://huggingface.co/blog/ibm-granite/granite-switch-4-1)
- Source URL: https://huggingface.co/blog/ibm-granite/granite-switch-4-1
- Ingested by Yomu: 2026-10-15T21:00:11.212Z

[Read original post](https://huggingface.co/blog/ibm-granite/granite-switch-4-1)