What happened
On October 15, 2026, the ibm-granite organization unveiled Granite Switch 4.1 models at 8B and 30B scales, which can shift efficiently between natural-language replies using a Decision adapter. The lineup includes Granite-Speech-5.0-Turbo-CTC, called the quickest ASR system on the OpenASR Leaderboard and embedded for transcription work. The post states that any user can assemble a personalized Switch model by combining adapters from Granite Libraries.
Why it matters
According to the post, the Granite Switch 4.1 models run in vllm through a plug-in, and the Decision adapter needs roughly the same GPU VRAM as the base model alone. Users can create custom Switch models from Granite Libraries adapters, and the post notes that a software development kit for training adapters within Mellea is planned for release.
Read the original at Hugging Face