Granite Switch 4.1

Not yet independently verified. Every detail here comes from IBM's own model cards and launch post, which date the release to 14 October 2026; no independent report was found. The Decision Index figure on the card is IBM's own measurement, so no score is stored. We will update this when it can be confirmed, and remove this note.

Granite Switch 4.1 is a modular language model from IBM, released on 14 October 2026 in two sizes built on the Granite 4.1 8B and 30B base models. Its idea is that one hosted endpoint can do several jobs a team would otherwise split across separate services: answer in natural language, make structured decisions, check content for safety risks and transcribe speech. The weights are published on Hugging Face under the Apache 2.0 licence.

What it does

Each checkpoint combines a base model with two activated-LoRA adapters and a speech encoder, and the adapters are switched on by control tokens in the chat template, so the calling code does not change. The decision adapter takes a state and a question with labelled options and returns a probability for each option in a single forward pass, which suits routing, triage and classification steps inside an agent. The Guardian Core adapter, taken from IBM's Granite Guardian library, flags harm, jailbreak attempts, profanity, violence, sexual content, social bias and unethical behaviour. Audio sent to the same endpoint is transcribed by the embedded Granite Speech 5.0 TurboCTC encoder of 470 million parameters, which handles English only. The 8B base model has a context of 128,000 tokens.

Who it is for

IBM aims it at developers who run their own inference with Hugging Face Transformers or vLLM and want fewer models to deploy. An open-source composer tool rebuilds the checkpoint, or composes a different one from IBM's adapter catalogue, and IBM's Mellea library wraps the Guardian adapters as direct operations; Mellea support for the decision adapter is planned but not yet available. Earlier preview checkpoints with the full Granite Libraries bundle, without audio, remain available in 3B, 8B and 30B sizes. IBM's card reports its own Decision Index results for the decision adapter; no score is stored here.

Key facts

Maker: IBM Released: 14 October 2026 Sizes: 8B and 30B base models, plus a 470M speech encoder Context: 128K tokens (8B base) Inputs: text and English audio; output text and option probabilities Licence: Apache 2.0 Runs on: Hugging Face Transformers and vLLM