Granite Switch 4.1
IBM released Granite-Switch-4.1-8B and -30B, which combine chat, a decision adapter, a Guardian safety adapter and fast speech recognition in one model served from one endpoint.
The models add a general Decision aLoRA, a specialized Guardian aLoRA and the Granite-Speech-5.0-Turbo-CTC ASR model to a base Granite model. The pieces use about the same GPU memory as the base model alone. The adapters are Activated LoRAs, so the KV cache from a base model generation can be reused for later adapter decisions.
- Date
- Thursday 15 October 2026
- Lab
- IBM
- Kind
- model
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Doom decisions per second (custom 3B Granite Switch model) | 60 - 100 compared with Jev's reported 10 decisions per second | company |
The license was not stated on the page read, so access is inferred from the Hugging Face model cards the post links to. The Guardian adapter is claimed to beat Jev 1.13 and other decision models on safety tasks, with no figures given.
Sources
This record was checked against its sources on 16 October 2026. How we check