2026-10-29 · ← News
IBM Patches Granite Switch to 4.1.1 by Swapping Adapters Without Retouching the Base Model
IBM released Granite Switch 4.1.1 as a backward-compatible patch that updates its 10-language speech recognition model and decision adapter while leaving the base model untouched. The release demonstrates that modular adapter architectures allow teams to ship incremental capability upgrades on a single endpoint without recomputing core weights or forcing user migrations.
The image could not be loaded.
Granite Switch 4.1.1 Updates Speech and Decision Modules Over an Untouched Base
Just over a week after introducing the modular Granite Switch 4.1 architecture, IBM released Granite Switch 4.1.1 on October 29, 2026. The update functions as a backward-compatible software patch that requires no migration from version 4.1. Instead of retraining the entire system, IBM left the base language model untouched, recomputed only two specialized components, and published three models, one collection, and an updated Speech-to-Decisions demo Space on Hugging Face.
The first updated component is the Granite Speech automatic speech recognition model, which transcribes audio across 10 languages and now outputs native punctuation and capitalization. The second component is a revised Decision adapter in the granitelib-core library that improves overall performance on the Jev Decision Index test set. Users can compose these cache-efficient adapters and deploy them through a single inference endpoint.
Modular Adapter Patching Cuts Recomputation and Transcript Formatting Steps
Standard model upgrades often require full weight updates, endpoint changes, and fresh validation across every downstream task. By isolating improvements inside modular adapters, IBM treats model maintenance like routine software patching, shortening release cycles and avoiding the compute cost of retraining base weights.
The speech recognition upgrade also simplifies production pipelines. Because the 10-language speech model generates punctuation and capitalization natively, transcripts feed directly into downstream decision adapters without a separate text-formatting pass.
Missing Benchmark Numbers and an Unreleased Adapter SDK Limit Independent Use
Despite pitching software-style modularity for outside developers, IBM has not yet released the tools required for third-party teams to build their own adapters. Developer tutorials and an Adapter SDK remain in development, meaning external users can currently only compose modules supplied by IBM.
The announcement also omits quantitative verification. IBM provides no concrete scores for the Jev Decision Index improvement, does not list the 10 supported speech languages or their transcription error rates, and shares no latency or compute cost figures for running composed adapters.
Upcoming Adapter SDK and Quantitative Benchmarks Will Test the Modular Workflow
The primary milestone to watch is the launch of the promised Adapter SDK and accompanying tutorials. Only when independent developers can train, version, and swap custom adapters on top of the frozen base model will Granite Switch prove its utility beyond internal IBM patch cycles.
Technical evaluators should also watch for detailed breakdowns of the Jev Decision Index and per-language transcription accuracy across the 10 supported languages. Independent latency benchmarks on the single-endpoint architecture will further clarify whether modular adapter composition holds up under production speech-to-decision workloads.
Lilith's verdict
An infrastructure engineer swaps two fresh adapter weights onto a running Granite Switch 4.1 endpoint and watches raw audio in 10 languages turn into punctuated transcripts for the decision module without touching the base model. Until IBM ships the promised Adapter SDK, however, outside teams can only install patches from the vendor catalog rather than forging custom modules of their own.