Realtime AI News
IBM's Granite Switch 4.1.1 updates adapters, not the base model
IBM released Granite Switch 4.1.1, a patch version that is fully backwards compatible with 4.1 and requires no migration, changing only a multilingual speech recognition model and the Decision adapter in granitelib-core. IBM says nothing beyond those two components was recomputed, leaving the base model untouched while it builds an Adapter SDK for third parties.
IBM published Granite Switch 4.1.1 on October 29 and described it as a patch release in the software sense: fully backwards compatible with 4.1, no migration, incremental improvements under the hood.
Two components changed. The updated Granite Speech ASR model handles multilingual speech-to-text across 10 languages with native punctuation and capitalisation, so transcripts can go straight to downstream models without a formatting pass. The Decision adapter in granitelib-core was also updated, with improved overall performance on the Jev Decision Index test set.
What IBM emphasises is the cost of shipping it. Apart from training the new ASR model and Decision adapter, nothing was recomputed and the base model is untouched. Developers compose the adapters they want and deploy.
That follows from the modular Switch architecture, in which a language model is configured with specialised, cache-efficient adapters so new expertise can be added on a single endpoint. Granite Switch 4.1, released on October 15, was built around exactly that pitch.
IBM argues the pattern extends beyond its own patch releases. The team is 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 they ship software.
The models named in the post are granite-speech-5.0-470m-turboctc for speech recognition, granite-switch-4.1-30b and granitelib-core-r1.0, alongside a Speech-to-Decisions demo space that runs live speech through recognition and decision-making on one Granite Switch model.
For enterprises, the appeal is governance cost. A model update usually means re-evaluation, redeployment and sometimes a fresh procurement conversation, while swapping an adapter against an unchanged base narrows the change surface to an auditable module, which is more workable for regulated industries that need to keep iterating.
The variable to watch is the Adapter SDK. If adapters can genuinely be versioned, tested and composed like software packages, capability upgrades start to resemble application development rather than a once-a-year generational swap.
Why it matters
Granite Switch breaks model updates into swappable modules, cutting the re-evaluation and redeployment cost that slows enterprise adoption. An open Adapter SDK could push model iteration toward normal software release cadence.
Nearby Updates
All10/29, 21:00
Deepgram buys Pipecat maker Daily, pledging the framework stays open source
Deepgram is acquiring Daily, the company behind the widely used open-source voice AI framework Pipecat, for an undisclosed amount. Both companies say Pipecat will remain open source, and Deepgram is also launching an open protocol called the Real-Time Context Standard developed with the Pipecat team.
10/29, 21:00
Amazon adds 100-plus shopping capabilities to Alexa before the holidays
Amazon has added more than 100 shopping-related capabilities to Alexa, spanning list creation and editing, subscription changes, delivery instruction updates and billing questions. The company says use of its agentic features has grown eightfold this year, with over 400 million consumers shopping through Alexa in the past 12 months.
10/29, 20:45
Adversa AI launches a security platform to block attacks on coding agents in real time
Adversa AI has launched a coding-agent security platform that it says blocks cyberattacks on AI agents in real time, according to a PR Newswire release dated October 29. The product targets a specific shift: once agents can read and write code and call tools, the attack surface moves from model output to real actions.
10/29, 20:44
Infisical raises $40M at a $600M valuation to secure the AI agent economy
Infisical has announced a $40 million round at a $600 million valuation, framing the raise around securing the AI agent economy, according to an AI Insider report dated October 29. The company sits in secrets and credential management, a layer under mounting pressure as agents connect to far more systems than human users ever did.