IBM is betting one hosted model can do it all. Its new Granite Switch 4.1 family blends a general-purpose decision adapter, a safety specialist and a speech engine billed as the fastest of its kind. Just how far that mix goes, and how quickly it switches gears, is where things get interesting.
IBM's Granite Switch 4.1 models combine a general-purpose Decision adapter, a specialized Guardian adapter, and the fastest ASR model on the OpenASR leaderboard into one model. They can move between natural-language responses, decision tasks, and speech-to-text, all from one hosted endpoint. The models are Granite-Switch-4.1-8B and Granite-Switch-4.1-30B.
The Decision adapter is new and handles general-purpose decision-making, such as the v1/systemone decisions the article cites. The Guardian aLoRA is specialized for safety-related decision tasks. The article says it outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on average across those tasks. The speech component is Granite-Speech-5.0-Turbo-CTC, which IBM describes as the fastest ASR model on the OpenASR Leaderboard. It is embedded in the model for transcription tasks.
All three technologies, along with the underlying base model, are composed into one model. The article says this takes up roughly the same amount of GPU VRAM as the base model alone, which means more capabilities with fewer GPUs. The Switch adapters are built as cache-efficient Activated LoRAs (aLoRAs). The KV cache created during a base model generation can be reused for any later adapter-based decision. IBM also cites its recent vLLM contributions for speeding up low-rank adapter inference, and a co-located ASR encoder. Together, these let a Granite Switch 4.1 model switch between a chat request, a decision request, and a spoken request quickly. IBM says the 8B model can transcribe audio, generate summaries, screen for harmful content, and test 15+ generic flags, all on a single endpoint in real time.
Decision models are not new for IBM. The earlier Granite Libraries release includes a repository of specialized “decision model” adapters and other task adapters that can be composed into custom Granite Switch models. IBM argues that many of its specialized granitelib adapters can outperform generalist models on specialized classification tasks. Because adapters share the base model's KV cache, adding specialists alongside a generalist decision adapter costs almost nothing. The result, according to IBM, is broad coverage from the generalist and higher accuracy from specialists on the tasks that matter.
IBM ran two demonstrations. In a “race,” Granite-Switch-4.1-30B (a frozen Granite base with four specialized aLoRAs) completed a real-world, five-step RAG workflow much faster than Jev-27B (a frozen Qwen base with one generic decision LoRA). In a second demo, a custom 3B Granite Switch model navigated the Doom game at 60 to 100 decisions per second, compared with Jev's reported 10 decisions per second. IBM notes that unlike Jev, this model can also narrate and respond to questions while it plays.
The released 8B and 30B models come “batteries included,” compiled with a predefined set of adapters. Users can also compose their own Granite Switch models by mixing adapters from Granite Libraries. IBM plans to release an Adapter SDK for training adapters in Mellea, along with tutorials and further Granite Switch support, so the community can add custom and third-party adapters.
Key Features at a Glance
Decision adapter: A new general-purpose adapter for decision-making, suited to broad coverage across tasks.
Guardian aLoRA: Specialized for safety-related decisions. IBM says it outperforms four named generalist decision models on average.
Granite-Speech-5.0-Turbo-CTC: Embedded speech transcription, described as the fastest ASR model on the OpenASR Leaderboard.
Shared GPU footprint: All components sit in one model at roughly the VRAM cost of the base model alone.
Reusable KV cache: Activated LoRAs reuse the base model's cache for later adapter decisions, which speeds up switching.
Composable design: Users can mix and match Granite Libraries adapters to build custom Granite Switch models.
Availability
The article does not give a release date or pricing. The Granite-Switch-4.1-8B and -30B models are on Hugging Face, and the Granite Switch 4.1 model cards are in an IBM Granite collection there. The project code is on GitHub, and a Speech-to-Decisions demo is available on Hugging Face Spaces.
