IBM's Granite organization published Granite Switch 4.1 on October 15, 2026, extending the modular Switch approach introduced earlier in the year. The release centers on Granite-Switch-4.1-8B and Granite-Switch-4.1-30B configurations that keep a frozen base model while attaching Activated LoRA adapters for decisioning and a co-located speech encoder for transcription.
According to the Hugging Face Blog post authored by IBM Granite researchers, the 4.1 builds include a general-purpose Decision aLoRA for tasks such as system-style decisions, a specialized Guardian aLoRA aimed at safety-related decision tasks, and Granite-Speech-5.0-Turbo-CTC, which the post describes as the fastest automatic speech recognition model on the OpenASR leaderboard at the time of writing. The adapters plus speech stack are composed so that the combined footprint stays close to the VRAM of the base model alone.
Cache reuse is the systems claim. Because Switch adapters are built as cache-efficient Activated LoRAs, a KV cache produced during base-model generation can be reused for subsequent adapter-based decisions. IBM also cites recent vLLM contributions that speed low-rank adapter inference. Together with the co-located ASR encoder, the company says the same endpoint can move from a chat request to a decision request to a spoken request without loading separate full models.
IBM positions Granite Libraries as the repository of specialized decision and task adapters that users can mix into custom Switch compositions rather than relying only on the batteries-included 4.1 bundles. The post contrasts specialized granitelib adapters with generalist decision models on specialized classification tasks and shows a workflow race narrative in which a frozen Granite base with multiple specialized aLoRAs completes a multi-step RAG path faster than a competing generalist decision setup. Those comparisons are vendor-presented demos, including a Doom-playing 3B Switch anecdote claiming 60 to 100 decisions per second versus a cited competitor figure of about 10.
A Speech-to-Decisions Hugging Face Space is offered as a live demo where granite-switch-4.1-8b transcribes audio, summarizes, screens for harmful content, and evaluates additional user-defined flags on one endpoint. IBM says it plans an Adapter SDK for training adapters in Mellea, with more tutorials to follow. The announcement does not publish full OpenASR score tables or independent replication of the Guardian average-outperformance claim versus named closed decision models.
Practically, that means an operations team can evaluate whether Guardian flags and ASR transcripts share the same serving process that already hosts chat completions, instead of stitching three separate model deployments. The post leaves open how multi-tenant hosting and rate limits behave when speech encoding and decision adapters run in the same request path.
For practitioners, Granite Switch 4.1 is best read as a packaging and systems release: one endpoint that multiplexes chat, safety decisions, and speech while preserving KV cache across adapter hops. Buyers should treat leaderboard and race claims as IBM-reported until third-party measurements appear, and note that composing custom adapters still depends on forthcoming SDK tooling.