IBM's Granite Switch 4.1 puts chat, decisions and speech in one model

The team behind IBM's Granite model family released Granite Switch 4.1 on October 15, 2026. The two new models, Granite-Switch-4.1-8B and Granite-Switch-4.1-30B, are designed to handle natural-language chat, general and specialized decision tasks, and speech transcription from a single hosted endpoint.

According to the developers, each model bundles three components: a new Decision aLoRA for general-purpose decisions, a Guardian aLoRA aimed at safety-related classification, and Granite-Speech-5.0-Turbo-CTC, which the team describes as the fastest speech recognition model on the OpenASR leaderboard. The team claims the Guardian adapter, on average, beats Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety decision tasks.

More capability on the same hardware

The developers say the combined model uses roughly the same GPU memory as the base model alone. Because the adapters are built as Activated LoRAs, the key-value cache produced during a base model generation can be reused by later adapter-based decisions, avoiding repeated computation. The team also credits its recent contributions to vLLM, which speed up low-rank adapter inference, and a co-located speech encoder.

In a comparison published by the team, Granite-Switch-4.1-30B, running a frozen Granite base with four specialized aLoRAs, finished a five-step retrieval-augmented generation workflow much faster than Jev-27B, which pairs a frozen Qwen base with one generic decision LoRA. The developers also argue that their specialized granitelib adapters can outperform generalist decision models on narrow classification tasks.

Build-your-own models

The released models ship with a predefined set of adapters, but users can assemble custom Granite Switch models from the Granite Libraries repository. The team plans an Adapter SDK for training adapters in Mellea, along with tutorials. As a demonstration, it says a custom 3-billion-parameter Granite Switch model played Doom at 60 to 100 decisions per second, compared with 10 per second reported for Jev, while also narrating its play.

The approach targets organizations that want moderation, transcription and conversational features without running separate models on separate GPUs. The performance figures come from the developers' own tests, and the team says more details and data will follow in the coming weeks.