Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

AI news & research

News ·

IBM's Granite Switch 4.1.1 adds speech recognition as a drop-in patch

IBM has shipped Granite Switch 4.1.1, a backwards-compatible patch to its adapter-based model system, just over a week after 4.1. It adds a 470M-parameter Granite Speech model that transcribes 10 languages with punctuation and a revised Decision adapter, while the 30B base model is untouched, so users update adapters rather than retraining or migrating.

IBM has shipped an update to its Granite Switch models that it numbers like a software patch, and the numbering is the point. Granite Switch 4.1.1 arrived just over a week after 4.1, is fully backwards compatible, and needs no migration. For companies that have to re-test and re-approve every model change, an update that leaves the main model alone is much cheaper to accept than a new release.

Granite Switch is IBM's modular design: one base model with a set of small adapters, each trained for a job, that can be switched on as needed from a single hosted endpoint. Version 4.1 came in 8B and 30B sizes and bundled a general decision adapter, a safety-screening adapter and a fast speech recognition model. Because the adapters are built to reuse the base model's cached work, IBM says the whole bundle takes roughly the same GPU memory as the base model alone, and the same endpoint can answer chat requests, yes-or-no decisions and transcription jobs.

The patch changes two parts. The speech component is now Granite Speech 5.0 TurboCTC, a 470M-parameter transcription model that has become multilingual and adds punctuation and capital letters, so transcripts can go straight to another model without a clean-up step. IBM's model card says the multilingual version also slightly improves average English word error rate, from 5.78% to 5.51%. The second change is a revised decision adapter that IBM says scores better on the Jev Decision Index test set. The 30B base model was not retrained.

This fits a wider pattern of small, specialised models wrapped around a stable core, and it suits the cautious enterprise buyers that our review of the Granite 4.2 family identifies as IBM's audience. Readers comparing transcription options should note that this TurboCTC model is a smaller, faster sibling of Granite Speech 5.0, which we review separately. IBM says an Adapter SDK and tutorials are coming so that other developers can build and version their own adapters.

There are gaps. IBM gives no numbers for the revised decision adapter, so its improvement is a claim, not a result anyone can check yet. The language count is also inconsistent: the release post says 10 languages, while the model card lists seven, English, French, German, Spanish, Portuguese, Italian and Japanese. Teams relying on a particular language should test it before upgrading. And the promise that adapters can be swapped without side effects will only be proven once third parties start shipping their own.

Our take: Treating models like versioned software, where you swap one adapter and leave the rest alone, is a practical idea for enterprises tired of full re-evaluations. IBM gives no scores for the revised Decision adapter, so its claimed improvement is unverified for now.

Our reviews:Granite Speech 5.0 2BGranite 4.2

Source: Hugging Face

More stories