IBM Granite Switch 4.1.1 Adds 10-Language Speech Updates
IBM Granite Switch 4.1.1 adds speech recognition across 10 languages and a stronger Decision adapter, with no base-model recomputation or migration required.
Summary
IBM released Granite Switch 4.1.1 on October 29, 2026, just over a week after Granite Switch 4.1 introduced a modular architecture for combining specialized, cache-efficient adapters on one endpoint. The patch is fully backward compatible with 4.1, requires no migration and leaves the base model unchanged.
The release updates Granite Speech ASR with speech-to-text across 10 languages, plus native punctuation and capitalization that eliminate a formatting pass before transcripts reach downstream models. It also updates the Decision adapter in granitelib-core, improving overall performance on the Jev Decision Index test set, although IBM provided no score. Only the new ASR model and Decision adapter required training, with nothing else recomputed. IBM is preparing tutorials and an Adapter SDK so developers and third-party providers can build, test, version and ship adapters like software. The new capabilities are available through the Speech-to-Decisions demo, but IBM gave no release date for the developer tooling.
Positives
- Granite Switch 4.1.1 remains fully backward compatible with version 4.1 and requires no migration.
- Granite Speech ASR now transcribes 10 languages with native punctuation and capitalization.
- The updated Decision adapter improves overall performance on the Jev Decision Index test set.
- IBM left the base model untouched, avoiding recomputation beyond training the new ASR model and Decision adapter.
- One endpoint can combine specialized, cache-efficient adapters and deploy their latest versions.
Risks & concerns
- IBM provided no numerical results for the Decision adapter's improvement on the Jev Decision Index.
- The Adapter SDK and tutorials remain in development, with no announced release date.
- The Granite Speech ASR model and Decision adapter still required training despite the unchanged base model.