← All model news

Granite Switch 4.1.1 patch ships with unchanged base, new speech model and Decision adapter

IBM released the Granite Switch 4.1.1 patch on October 29, 2026, saying it is fully backward compatible with 4.1 and needs no migration. It updates the Granite Speech ASR model, which covers 10 languages and natively outputs punctuation and casing, and the Decision adapter in granitelib-core; the base model was not recomputed.

Published Updated
Cover from the original report — Granite Switch 4.1.1 patch ships with unchanged base, new speech model and Decision adapter
Cover from the original reportOriginal report ↗

What changed

According to a blog post published by the IBM Granite team on Hugging Face on October 29, 2026, Granite Switch 4.1.1 is a patch release to 4.1, arriving a little more than a week after 4.1. The team explains it using software versioning conventions: it is fully backward compatible with 4.1, requires no migration, and all changes are underneath. The post lists two updates. The first is the Granite Speech ASR model granite-speech-5.0-470m-turboctc, which covers 10 languages and natively outputs punctuation and casing. The team says transcripts can therefore be fed directly into downstream models without another formatting pass. The second is the Decision adapter in granitelib-core, which the team says improves overall performance on the Jev Decision Index test set. Notably, the post uses different labels for the two items: the speech item is called a 'new ASR model,' and only the decision item is called an adapter. The post does not say whether the speech capability is distributed as a hot-swappable adapter or deployed as a standalone model. The team attributes the speed of this release to not recomputing the base: apart from training the new ASR model and the Decision adapter, the base model was unchanged, and users can combine the adapters they need and then deploy. The post also mentions that tutorials and an Adapter SDK are in progress so that developers and third-party providers can build their own adapters, but it gives no timeline, interface scope or distribution method. The blog also points to a Speech-to-Decisions demo. The post gives no specific benchmark scores, no list of the 10 languages, no license terms, and no latency, throughput or memory usage data.

What it means for developers

The signal in this release is its form, not its scores. With the base fixed and only capability components swapped, the unit of update has moved from the whole model to replaceable parts. For teams already integrated with 4.1, interface-level migration cost is close to zero, so the risk shifts to version drift: if the deployment pipeline pulls the 'latest adapter' rather than a pinned revision, model behavior can change without any code change, and the patch digit in the version number itself signals that these small-step replacements will continue. Reading 'backward compatible' as 'outputs unchanged' would be a misreading; the compatibility in the post refers to the interface and composition relationship with 4.1, not to verbatim transcription or decision results. The direct change in the speech pipeline is the possible removal of one post-processing step. If native punctuation and casing work as described, the previous 'ASR output then format' stage could save one call and one maintenance cost; but the shape of the text sent downstream also changes, and prompts and parsing logic that rely on fixed punctuation, all-lowercase text or specific segmentation assumptions need to be rechecked. Another precondition is language coverage: the post says only 10 languages and does not list them, so whether Chinese or other business languages are included needs to be confirmed independently. On the Decision adapter, 'overall improvement on the Jev Decision Index' is the publisher's own account on a single test set; an aggregate metric improvement does not mean the same direction holds for your own option counts, input lengths and domain distribution. In addition, the Adapter SDK and tutorials for third-party self-built adapters are still planned, and cannot now be treated as an available deployment option for planning.

Integration and evaluation

The following are conditional recommendations and analysis, not reported facts. First, pin the base and adapters separately by version in deployment, rather than following the latest as a whole. The basis is that 4.1.1 demonstrates exactly the 'base unchanged, components replaced' case. The choice is whether to lock the Decision adapter and speech model by revision or content digest in the deployment configuration, or to continue resolving 'latest.' Load conditions that would trigger a switch are: the endpoint is shared by multiple products, CI automatically pulls the latest weights, or output stability is a hard constraint downstream. A testable criterion: run your own holdout set once on the old components and once on the new, and switch only if the differences in decision accuracy and transcription quality metrics fall within a pre-set tolerance; if they exceed it, keep the old version and log the differing samples. Second, evaluate whether to remove the formatting step after speech. The basis is the new ASR model's claim to natively output punctuation and casing. The choice is whether to keep the two-stage 'ASR plus formatting' pipeline or move to ASR only. Preconditions are that the languages in your business audio are confirmed to be among the 10, and that the downstream task actually depends on punctuation or casing; if neither holds, there is no need to change the existing pipeline. A testable criterion: run the same batch of real business audio through both paths and compare downstream task accuracy and parse failure rate; remove formatting only if the ASR-only path is no worse than the two-stage path. Third, evaluate the Decision adapter by business slice rather than directly accepting the overall conclusion. If your option counts and input length distribution differ markedly from the Jev Decision Index, split your own historical decision samples into several slices and measure accuracy separately to see whether the improvement is concentrated in slices similar to your traffic; adopt it only if similar slices also improve, otherwise treat it as a neutral update.

Limits and open questions

The blog post gives no specific scores or test set composition for the Jev Decision Index, does not list which 10 languages are covered, and provides no latency, throughput or memory usage data. Therefore 'overall performance improvement' and 'native punctuation' cannot be verified outside the post and should be treated only as the publisher's account. The post does not state the license terms for the base or the components in this release; downloadable weights do not equal a commercial license, and specific rights must be based on the corresponding model card. The Adapter SDK and tutorials are described only as in progress, with no timeline, so a third-party self-built adapter route cannot be planned on that basis. Granite Switch 4.1.1, along with the updated Granite Speech ASR model and Decision adapter, are not currently listed on KimiSeek. The availability described in this article refers to IBM's release channel on Hugging Face and does not mean this site provides the model. For information on how models listed on this site are billed, please see the pricing page.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

Granite Switch 4.1.1——像更新软件一样更新模型

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.