One Model, Three Functions: IBM Launches Granite Switch 4.1 in 8B and 30B

One Model, Three Functions: IBM Launches Granite Switch 4.1 in 8B and 30B

IBM has opened up two modular language models under Apache 2.0 that combine a decision adapter, a safety adapter and a built-in speech recognizer on a single base model – reportedly without materially extra GPU memory. But all of the performance figures are, for now, IBM's own, and the company's own comparisons show that the specialist adapters lose to the generalist models on several benchmarks.

IBM launched Granite-Switch-4.1-8B and Granite-Switch-4.1-30B on October 15, 2026, with model cards on Hugging Face dated October 14 (Unite.AI). The models are licensed under Apache 2.0, have a 128,000-token context window, and support both Hugging Face Transformers and vLLM as serving backends. The core of the launch is not a new base model, but a composition: one base model equipped with three add-on components that together are meant to serve chat, decision evaluation and speech-to-text from a single endpoint (IBM).

How the composition works

The Granite Switch project is, according to IBM, a modular alternative to traditional model development, built on Activated LoRA (aLoRA) – lightweight adapters that can be activated on a fully trained base model. The Granite-Switch-4.1 models include three such components (IBM):

1. A Decision aLoRA for general decision evaluation. According to Unite.AI's review of the model cards, the adapter produces a softmax distribution over labeled options in a single forward pass – a classification directly, without generating free-form text.
2. A Guardian aLoRA, a specialized safety adapter that IBM claims beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash "on average" on safety-related decision tasks.
3. Granite-Speech-5.0-Turbo-CTC, a speech-to-text model that IBM describes as the fastest ASR model on the OpenASR Leaderboard, built in for transcription tasks.

"All three technologies (along with the underlying base model) are composed into one model, which occupies roughly the same VRAM on a GPU as the base model alone would. That means more capabilities with fewer GPUs," IBM writes in the Hugging Face blog post. IBM adds that aLoRA cache reuse, along with kernel contributions in vLLM, is supposed to make switching between chat, decision and speech requests fast – but the company provides no concrete latency or switching figures, so the claim remains qualitative.

The example IBM itself highlights is the 8B model: transcribe audio, generate summaries, screen for harmful content and evaluate over 15 generic flags – all on one endpoint, in real time (IBM). For a company that today has to run separate services for transcription, moderation and decision logic, this is – if the claims hold – the practical value proposition.

The numbers IBM itself has published

The decision adapter was measured on the public "fast" portion of the Decision Index benchmark, a subset of 1,068 rows. The figures, which Unite.AI reproduces from the model cards:

The 30B model is strongest on tool use at 76.5, and weakest on the Arts category for both sizes (18.0 and 26.4). These numbers are reported by IBM and are not independently verified.

Where the specialist claim hits pushback

IBM argues that specialized adapters often beat generalists on narrow classification tasks: "While Jev and other closed decision models focus on generality, we have found that many of our specialized granite-lib adapters can outperform the generalists on specialized classification tasks," the company writes (IBM). IBM also claims that a Granite-Switch-4.1-30B configuration completed a five-step RAG workflow faster than Jev-27B.

But IBM's own comparison against the generalist models also shows the opposite picture: the best Granite adapter came out behind on LM-AggreFact (by 2.6 points), InfoBench-GPT4 (3.2 points) and InfoBench-Human (0.8 points) (Unite.AI). The specialist claim thus applies primarily to narrow classification tasks – not decision evaluation in general. Readers evaluating the model should note that these are IBM's own benchmark figures, and that no independent evaluation of Granite Switch 4.1 exists in the source material.

What remains to be verified

Several questions are open after the launch:

• Switching speed: IBM describes switching between tasks as very fast, but provides no figures. Without quantified latency measurements, the claim remains qualitative.
• Hosting and pricing: It is unclear from the available sources whether IBM offers a hosting service for the model, and at what price.
• Audio language support: According to Unite.AI, the speech-to-text component is described as English-only. For users outside English-speaking markets, this is a significant limitation.
• The Guardian comparison: The claim that the Guardian adapter beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash is unspecified as to which benchmarks it is based on, and who ran them.

Why it matters now

Granite Switch 4.1 arrives at a time when several players are launching dedicated decision models, while IBM positions adapter-based specialization as an alternative: one modular model where specialization happens through swappable adapters instead of separate deployments. If the VRAM claim holds up in practice, it lowers the economic threshold for running multiple AI functions at once – particularly for organizations with limited GPU capacity.

But with open weights under Apache 2.0 comes the opportunity for independent parties to verify – or disprove – IBM's numbers. The model cards on Hugging Face make it possible for third parties to run Decision Index and the comparison benchmarks themselves. Until that happens, the central performance claims around Granite Switch 4.1 remain IBM's own.