Eigen RadarAI
Archive
Analysis

Granite Switch 4.1 combines chat and decisions with switchable adapters

IBM has detailed two modular Granite Switch 4.1 models that combine conversation, structured decisions and speech transcription. Specialized adapters reuse the base model’s context rather than requiring a separate model for each task. The October 15 announcement follows model files dated October 14. Speech support is limited to English, and developers can compose their own adapter combinations.

Artificial Intelligence··Midday
Different circuit cards sit in separate sockets on a shared copper-traced backplane in a bright workshop.

Adapters share one model endpoint

IBM, the technology company behind the Granite language models, detailed its modular Granite Switch 4.1 configuration on October 15. Two models combine conversation, structured decisions, safety screening and speech transcription in one endpoint. Their base models contain 8 billion and 30 billion parameters. The model cards separately date the files’ release to October 14. The announcement develops the Granite Switch project introduced the previous spring, with predefined combinations of components that can also be assembled differently.[1], [2]

The supplied configurations attach a general-purpose Decision adapter and a Guardian safety adapter to the underlying language model, alongside a speech encoder. An adapter adds behavior for a particular task. Developers can select other specialized components from Granite Libraries, the project’s collection of task adapters, to compose a custom Switch model instead of using the supplied combination.[1]

A selected adapter reads the base model’s context

Callers select an adapter by naming it in the chat template. A switching layer then directs a span of tokens, the units of text processed by the model, through that adapter across the decoder layers. Each adapter reads the base model’s attention cache, which retains previously processed context, without reading another adapter’s internal state. The adapters are developed independently rather than trained jointly.[2]

IBM says its Activated LoRA adapters can reuse the cache created during ordinary text generation for subsequent decisions. Its speech demonstration transcribes incoming audio, produces a summary and screens for harmful content within one endpoint. The same demonstration also checks flags defined by the user. The hosted combination brings those steps together while preserving separate specialized components.[1]

Transcription stays limited to English

Speech transcription supports only English and produces lowercase text without punctuation. The transcript becomes part of the conversation sent to the language model, so a transcription error can alter a later answer. Guardian does not replace application-level safety testing. The packages also omit some factuality and policy adapters. An adapter-training toolkit for Mellea, IBM’s application layer, remains a company plan rather than a delivered part of this announcement.[2]

References

  1. News sourceIBM GraniteIBM details modular Granite Switch 4.1 models with switchable adapters↩1↩2↩3
  2. News sourceUnite.AIIBM announces modular Granite Switch 4.1 models in two sizes↩1↩2↩3
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →