IBM ships Granite Switch 4.1.1, a patch with multilingual speech input and new decision weights

What IBM released

IBM published Granite Switch 4.1.1 on 29 October 2026 under the headline "Update models like you update software". The authors, Kate Soule, Luis A. Lastras, Subhajit Chaudhury and Abraham Daniels, call it a patch release in the software sense: fully backwards compatible with Granite Switch 4.1, no migration, and incremental improvements under the hood. Version 4.1 came out on 14 October.

The 8B model card and the 30B model card list two changes. The Decision activated LoRA adapter ships with updated weights, while its prompt format, invocation and readout stay the same, so existing integrations work as before. The embedded Granite Speech 5.0 TurboCTC encoder, 470 million parameters, now transcribes English, French, German, Spanish, Portuguese, Italian and Japanese and returns punctuated, capitalized text. In 4.1 it handled English only. The Guardian Core safety adapter is unchanged.

How the update was built

IBM's main point is the process rather than the features. The blog says that apart from training the new speech model and the new Decision adapter, nothing was recomputed and the Granite 4.1 base model was not touched. The new checkpoint was composed from the latest adapters with version 0.1.1 of the open-source granite-switch composer. The main branch of each Hugging Face repository holds the latest revision, and every release is tagged, so a team that wants 4.1 behavior can pin the revision granite-switch-4.1-8b in Transformers or vLLM.

On IBM's numbers, the updated Decision adapter scores 46.72 overall on the public Jev Decision Index v0.3 set for the 8B checkpoint and 51.33 for the 30B one. The 4.1 cards reported 42.9 and 50.1 on a fast subset of the same index, so the two sets of figures are not a like-for-like comparison. For speech, the encoder's own card says the multilingual update lowered average English word error rate from 5.78% to 5.51% on Open ASR leaderboard benchmarks.

Why it matters for open models

Most open-weight model families ship improvements as new full checkpoints that need fresh evaluation and redeployment. Granite Switch treats a model more like a package with versioned parts: the base stays fixed, and adapters or encoders can be replaced and tagged. If that holds up, operators can take a narrow fix, such as a better classifier or speech front end, without retesting the whole model. IBM says it is working on tutorials and an Adapter SDK so that developers and third-party providers can build and ship their own adapters the same way. More open-weight releases are tracked under Open Source AI and in the model release tracker.

What to watch next

The blog text says the speech model covers 10 languages, while both model cards and the encoder card list seven, so the exact language count should be checked against the cards before deployment. The cards also note that the speech encoder and the language model are not trained jointly, so a mistranscription can change what the model or an adapter responds to. No date is given for the Adapter SDK or for Mellea support of the Decision adapter, which the cards describe as planned.

Questions people ask

Do I need to change my code to use Granite Switch 4.1.1?

No, according to IBM. The patch is backwards compatible with 4.1, and the Decision adapter keeps the same prompt format, invocation and readout.

Which languages does Granite Switch 4.1.1 transcribe?

The model cards list English, French, German, Spanish, Portuguese, Italian and Japanese. IBM's blog post mentions 10 languages, which does not match the cards.

How do I keep using Granite Switch 4.1?

Pin the tagged revision, for example revision granite-switch-4.1-8b, when loading the model in Transformers or starting vLLM.

Sources

1. Granite Switch 4.1.1: Update models like you update software (IBM on Hugging Face, 29 Oct 2026)
2. Granite Switch 4.1 8B model card (revision 4.1.1) (IBM on Hugging Face, 27 Oct 2026)
3. Granite Switch 4.1 30B model card (revision 4.1.1) (IBM on Hugging Face, 27 Oct 2026)
4. Granite Speech 5.0 470M TurboCTC model card (IBM on Hugging Face, 21 Oct 2026)