IBM's Granite team has announced Granite Switch 4.1 in a post on Hugging Face dated October 15, 2026. It adds two models, Granite-Switch-4.1-8B and Granite-Switch-4.1-30B. Each one switches between natural-language answers, decision tasks and speech-to-text, all served from a single hosted endpoint.
What is in the package
The Granite Switch project, introduced last spring, builds a model from a base model plus small add-on adapters. The 4.1 models ship with three additions:
- A Decision adapter for general-purpose decisions.
- A Guardian adapter for safety-related decisions. IBM says it outperforms several other decision models on average on those tasks.
- Granite-Speech-5.0-Turbo-CTC, which IBM describes as the fastest speech recognition model on the OpenASR leaderboard.
All of this is composed into one model that, according to IBM, uses about the same GPU VRAM as the base model by itself. The adapters are built as Activated LoRAs, so the KV cache from a base-model generation can be reused for later adapter-based decisions. IBM adds that it has contributed vLLM changes to speed up low-rank adapter inference.
The demo and the comparison
IBM points to a Hugging Face space in which granite-switch-4.1-8b transcribes audio, writes summaries, screens for harmful content and tests 15 other user-defined flags on a single endpoint in real time. In a separate comparison, IBM says Granite-Switch-4.1-30B, with four specialised adapters, finished a real-world five-step RAG workflow faster than a rival setup using a 27B model with one generic decision adapter. These are IBM's own results. The post says more details and data will follow in the coming weeks.
Build your own
The two released models come with a preselected set of adapters, but IBM says anyone can compose their own Granite Switch model from adapters in its Granite Libraries. It argues that specialised adapters can beat generalist decision models on narrow classification tasks, and that adding them costs little because they share the base model's cache. IBM plans an Adapter SDK for training adapters in Mellea, with tutorials, but gives no date.
As a stress test, IBM built a custom 3B Granite Switch model that plays Doom. It reports 60 to 100 decisions per second, against 10 per second that a competing system has reported, and says its model can narrate and answer questions while playing.
What is not stated
The post gives no pricing, hosting plans or licence terms. It points readers to the model cards, a GitHub repository and a Colab notebook comparing aLoRA and LoRA.