IBM announced Granite Switch 4.1.1 on October 29, 2026, updating speech recognition and decision-making components while leaving the base language model untouched.
October 29 was the announcement, not the first checkpoint date
The primary evidence separates three milestones. IBM’s 30B model card dates the initial Granite Switch 4.1 revision to October 14, 2026; its introductory article was published on October 15. The same card dates revision 4.1.1 to October 27, preceding the patch announcement on October 29. The initial release therefore predates the requested October 20–November 1 reporting window; the patch falls inside it.
This is not merely an API-access announcement: IBM’s repository exposes downloadable Safetensors weights. Conversely, examples of an OpenAI-compatible interface on a locally served endpoint do not establish a new IBM-managed API launch. The original article already described hosted-endpoint operation. Separately named preview checkpoints carry a different, broader adapter bundle without audio; they should not be conflated with the speech-enabled release.
The patch changes two components—and exposes a language discrepancy
IBM identifies an updated speech recognizer and revised Decision adapter in granitelib-core. It reports improved overall performance on the Jev Decision Index test set, but that claim is not evidence of better results on a particular company’s ticket-routing or approval tasks. IBM also says the base model was not recomputed.
The language count deserves caution. The announcement says speech-to-text spans 10 languages. However, the linked Granite Speech 5.0 TurboCTC card documents seven: English, French, German, Spanish, Portuguese, Italian and Japanese. That card dates its multilingual update, including punctuation and capitalization, to October 21, 2026. This was an existing component update incorporated into Switch, not speech technology first introduced on October 29.
For procurement and rollout planning, use those seven documented languages as the supported baseline unless IBM supplies checkpoint-specific confirmation of additional coverage. That is our conservative reading of conflicting primary documentation, not a claim that the announcement has been formally corrected.
Why adapter-level updates matter to business agents
The Core Library’s Decision capability selects among supplied options and returns probabilities. This makes a support-routing scenario a useful evaluation target: transcribe a customer’s request, then select the responsible team from a bounded list. The adapter’s constrained answer space is useful, but it does not establish that the selected team is correct.
Our operational interpretation is that smaller update boundaries can make investigations more focused. If a routing outcome changes after this patch, compare the transcript first, then the decision made from identical text. That separates an upstream recognition change from a downstream classification change instead of treating the entire agent as one opaque replacement.
The audio implementation remains a speech-to-text cascade, explicitly labeled alpha in the project documentation: recognized text becomes ordinary input to the language model. It is not jointly trained end-to-end audio understanding. A polished transcript can therefore still carry a recognition mistake into the next decision.
What to validate before adopting 4.1.1
For a controlled rollout, we recommend a small paired evaluation rather than assuming the patch number guarantees equivalent business behavior:
- Keep a rollback target. Record the exact model revision and component configuration used in each evaluation.
- Replay representative audio. Include customer names, reference numbers, accents and background noise from the intended workflow.
- Separate recognition from routing. Compare both revisions using the same verified transcript as well as the original audio.
- Measure consequential errors. Review incorrect destinations, missed escalations and human overrides—not transcript appearance alone.
IBM’s October 29 post describes tutorials and an Adapter SDK as work in progress, not shipped developer tooling. The immediate opportunity is the verified patch and its downloadable artifacts; future extensibility should remain a separate planning assumption.