Glonce

IBM ships Granite Switch 4.1.1 with updated speech and Decision models

modelsThe Glonce editorsHugging Face🔥55Glonce rating55Impact35Novelty40Relevance100Surprise25Players60Credibility85
IBM ships Granite Switch 4.1.1 with updated speech and Decision models

IBM has released Granite Switch 4.1.1, a patch release of its modular Switch architecture. It arrives just over a week after Granite Switch 4.1, which IBM says showed that the Switch architecture lets you configure a language model with specialized, cache-efficient adapters, adding new expertise on a single endpoint.

The authors stress that the version number means what it means in software: 4.1.1 is fully backwards compatible with 4.1, needs no migration, and brings incremental improvements under the hood.

Two components are updated. The first is the Granite Speech ASR model, which offers multilingual speech-to-text across 10 languages with native punctuation and capitalization, so transcripts are ready for downstream models without a formatting pass. The second is the Decision adapter in granitelib-core, which IBM describes as having improved overall performance on the Jev Decision Index test set. No figures are given for that improvement.

The authors say the notable part is not the features themselves but how quickly and efficiently their teams could ship the update. Apart from training the new ASR model and Decision adapter, nothing was recomputed: the base model is untouched. Users compose the latest adapters they want and deploy.

IBM presents this as a pattern that can go well beyond patch releases to an IBM-provided model. It is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters, iterate on versions, test features, and ship updates the way software is shipped. The post ends with "More soon" and points readers to the latest models and to the Speech-to-Decisions demo, which is powered by the new multilingual and punctuation capabilities.

Key facts

  • Granite Switch 4.1.1 is a patch release, issued just over a week after 4.1, and is fully backwards compatible with it with no migration.
  • The updated Granite Speech ASR model does multilingual speech-to-text across 10 languages with native punctuation and capitalization.
  • The updated Decision adapter in granitelib-core is described as performing better overall on the Jev Decision Index test set; no numbers are given.
  • Apart from training the new ASR model and Decision adapter, nothing was recomputed: the base model is untouched.
  • IBM is working on tutorials and an Adapter SDK so developers and third-party providers can build their own adapters.

Why it matters

The release itself is small. The point IBM makes is about process: because Switch models are built from adapters on top of an untouched base model, only the new ASR model and Decision adapter had to be trained, and the update shipped about a week after 4.1. IBM frames this as updating models the way you update software, with patch version numbers that mean what they mean in software.

Who it affects

Teams already running Granite Switch 4.1 are the direct audience, since 4.1.1 is backwards compatible and needs no migration. Developers who want speech transcripts with punctuation and capitalization already in place, in any of 10 languages, are also covered. IBM also names developers and third-party providers as the future audience for the Adapter SDK, so they can build their own adapters.

How to use it

According to IBM, you compose the latest adapters you want and deploy. The post invites readers to test the latest models and to try the Speech-to-Decisions demo, which showcases the new multilingual and punctuation capabilities. The tutorials and Adapter SDK are still in progress, and the post says only "More soon".

How solid is it

This is a first-party IBM Granite blog post, so every claim comes from the vendor. The post names the updated components and the compatibility promise clearly. The claim of better Decision adapter performance rests on the Jev Decision Index test set, with no figures. The ASR model is described by its features, not by accuracy numbers.

Risks and caveats

The 10 supported languages are not listed, and no ASR accuracy figures such as word error rate are given. The Decision adapter improvement is stated without numbers. No release date or timeline is given for the tutorials and Adapter SDK beyond "More soon", so third-party adapter building is a plan, not something available now. Model sizes, licence and download links are not stated in the text.

“The version number means what it means in software: fully backwards compatible with 4.1, no migration, incremental improvements under the hood.”

— IBM Granite team, Hugging Face blog