IBM introduced Granite Switch 4.1 on October 15, 2026, combining chat, decision-making and speech-to-text capabilities in hosted 8B and 30B models that switch tasks through one endpoint. The company says the models include a general-purpose Decision aLoRA, a Guardian aLoRA for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC, which IBM calls the fastest ASR model on the OpenASR leaderboard.

IBM says the three technologies and the base model use approximately the same GPU VRAM as the base model alone. The Switch adapters are cache-efficient Activated LoRAs, allowing the KV cache from a base-model generation to be reused for later adapter-based decisions. IBM also cites its vLLM contributions to faster low-rank adapter inference and a co-located ASR encoder as supporting quick movement among chat, decision and spoken requests.

IBM's Speech-to-Decisions demo transcribes audio, generates summaries, screens for harmful content and checks 15 other user-defined flags in real time on one endpoint. The company says specialized Granite Libraries adapters can outperform generalist decision models on specialized classification tasks; sharing the base model's KV cache makes adding specialists alongside a generalist decision adapter cost almost nothing.

In a five-step RAG workflow comparison, Granite-Switch-4.1-30B, using a frozen Granite base and four specialized aLoRAs, completed the workflow much faster than Jev-27B, which used a frozen Qwen base and one generic decision LoRA. IBM says users can compose custom Switch models from Granite Libraries adapters and plans an Adapter SDK for training adapters in Mellea. As an example, its custom 3B-parameter Granite Switch model made 60–100 decisions per second while navigating Doom, compared with Jev's reported 10; unlike Jev, it could also narrate and answer questions while playing.

IBM introduced Granite Switch 4.1 on October 15, 2026, combining chat, decision-making and speech-to-text capabilities in hosted 8B and 30B models that switch tasks through one endpoint. The company says the models include a general-purpose Decision aLoRA, a Guardian aLoRA for safety-related decisions, and Granite-Speech-5.0-Turbo-CTC, which IBM calls the fastest ASR model on the OpenASR leaderboard.

IBM says the three technologies and the base model use approximately the same GPU VRAM as the base model alone. The Switch adapters are cache-efficient Activated LoRAs, allowing the KV cache from a base-model generation to be reused for later adapter-based decisions. IBM also cites its vLLM contributions to faster low-rank adapter inference and a co-located ASR encoder as supporting quick movement among chat, decision and spoken requests.

IBM's Speech-to-Decisions demo transcribes audio, generates summaries, screens for harmful content and checks 15 other user-defined flags in real time on one endpoint. The company says specialized Granite Libraries adapters can outperform generalist decision models on specialized classification tasks; sharing the base model's KV cache makes adding specialists alongside a generalist decision adapter cost almost nothing.

In a five-step RAG workflow comparison, Granite-Switch-4.1-30B, using a frozen Granite base and four specialized aLoRAs, completed the workflow much faster than Jev-27B, which used a frozen Qwen base and one generic decision LoRA. IBM says users can compose custom Switch models from Granite Libraries adapters and plans an Adapter SDK for training adapters in Mellea. As an example, its custom 3B-parameter Granite Switch model made 60–100 decisions per second while navigating Doom, compared with Jev's reported 10; unlike Jev, it could also narrate and answer questions while playing.