TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
IBM’s Granite team introduced Granite Switch 4.1 models that combine natural-language chat, decision-making adapters and speech transcription through one hosted endpoint. The company says the adapters use about the same GPU memory as the base model and reports advantages on selected tasks, but detailed independent comparisons and deployment results are not provided in the source report.
IBM’s Granite team has introduced Granite Switch 4.1, a family of hosted models designed to handle chat, decision tasks and speech transcription through one endpoint. The release combines Granite base models with decision adapters and an embedded speech recognition model, which IBM says adds capabilities while using approximately the same GPU memory as the base model alone.
The release includes Granite-Switch-4.1-8B and Granite-Switch-4.1-30B. IBM says both are configured with a new Decision aLoRA for general-purpose decisions, a specialized Guardian aLoRA for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC for transcription. The company describes the speech model as the fastest ASR model on the OpenASR leaderboard; the report does not provide the leaderboard results or measurement details.
IBM says the adapters and base model are composed into one model, taking approximately the same GPU VRAM as the base model by itself. The company also says its Activated LoRAs can reuse the base model’s key-value cache for later adapter-based decisions. Alongside a co-located speech encoder and recent vLLM work on low-rank adapter inference, that design is intended to let a model move between chat, decisions and audio requests without loading separate full models for each task.
A linked Speech-to-Decisions demo uses the 8B model to transcribe audio, summarize it, screen for harmful content and check 15 additional user-defined flags. IBM says the demo runs in real time on one endpoint. The report also describes a five-step retrieval-augmented generation workflow comparison between Granite-Switch-4.1-30B, configured with four specialized adapters, and Jev-27B, configured with one generic decision adapter. IBM says Granite completed the workflow faster, but gives no timings in the supplied material.
One Endpoint for Three Workloads
Combining language generation, classification-style decisions and transcription behind one endpoint could simplify systems that process spoken requests and then need to summarize or screen their contents. If IBM’s memory and cache-reuse claims hold in deployment, teams may be able to add specialized tasks without keeping a separate full model for each one. That could matter for GPU capacity and latency, particularly in workflows that apply several checks to the same input.
The release also presents a modular alternative to relying on one general-purpose decision model. IBM argues that task-specific adapters can improve performance on specialized classification while sharing the base model’s cache. The practical value depends on results for each workload, however; the source report does not include enough measurements to establish how the approach compares across different hardware, traffic patterns or evaluation sets.
As an affiliate, we earn on qualifying purchases.
Granite’s Adapter-Based Approach
The Granite Switch project was introduced the previous spring, according to IBM, as a modular approach to model development. The new 4.1 release packages a set of adapters with Granite base models, while the company says users can build custom combinations from adapters in Granite Libraries. That library contains specialized decision and task adapters intended to be composed with a general-purpose model.
IBM says specialists may outperform generalist decision models on narrow classification tasks. It cites comparisons with Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash for the Guardian adapter, and a separate workflow race against Jev-27B. These are company-reported claims. The supplied report does not include evaluation data, test procedures or independent confirmation, so readers cannot assess the comparisons from the announcement alone.
“All three of these technologies (along with the underlying base model) are composed into one model, taking up approximately the same amount of VRAM on a GPU as the base model would by itself.”
— IBM Granite team
As an affiliate, we earn on qualifying purchases.
Performance Evidence Still Pending
The source report does not state its publication date, and its claims about leaderboard speed, safety-task performance, GPU memory use and workflow speed are not accompanied by detailed measurements in the supplied material. The specific evaluation sets, hardware configurations, latency figures and comparison procedures are not described. It is also unclear how the hosted endpoint is priced, what availability or service limits apply, and whether the reported demo results generalize to production workloads.
IBM’s comparisons with named decision models are presented by the company; the report does not cite independent evaluations. The 15 user-defined flags in the demo are described, but their contents and accuracy are not specified. Those details will affect how developers can judge the model for safety screening or other consequential decisions.
As an affiliate, we earn on qualifying purchases.
More Adapter Details Expected
IBM says it plans to publish more details and data on Granite Switch in the coming weeks. The team also plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and further Granite Switch support. Developers can consult the 4.1 model cards, try the Speech-to-Decisions demo and use the Granite Switch repository and Granite Libraries to explore or compose adapter configurations.
The next useful evidence will be reproducible performance results that specify hardware, task data and comparison methods, alongside deployment details for the hosted endpoint. Those materials would help developers assess whether the combined model’s speed and memory characteristics hold for their own workloads.
speech to text transcription device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Granite Switch 4.1?
It is a set of 8B and 30B Granite models configured with adapters for chat-related workflows, general and safety-related decisions, and speech transcription through one hosted endpoint.
Which speech model does the release include?
IBM says the models include Granite-Speech-5.0-Turbo-CTC for transcription and describes it as the fastest ASR model on the OpenASR leaderboard. The supplied report does not give the underlying leaderboard measurements.
Does IBM show that the models outperform competitors?
The report says IBM’s Guardian adapter performs better on average than four named systems on safety-related decision tasks, and that Granite Switch 4.1-30B completed a five-step workflow faster than Jev-27B. These are IBM-reported comparisons; the supplied material does not include detailed results or independent verification.
Can developers create custom Granite Switch models?
IBM says users can combine adapters from Granite Libraries to build custom models. The team also plans an Adapter SDK for training adapters in Mellea, plus tutorials and additional support.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
