All stories
Granite Switch 4.1 Launches Developing

IBM packs chat, decisions and English speech into one Granite Switch endpoint

IBM's 8B and 30B checkpoints combine a chat base, Decision and Guardian adapters, and an English speech encoder, with cache reuse as the efficiency pitch.

By AITrending2 min read
Sources (3)

Why now IBM published its English announcement on October 15. The two model cards give October 14 as their release date; the announcement and updated documentation explain the combined speech and decision configuration rather than a first launch of the Switch project.
IBM's conceptual Granite Switch illustration with a large base-model block and three smaller adapter blocks.
IBM's conceptual adapter illustration from its English announcement. The blocks explain modular composition; they are not measured memory sizes or benchmark results. Source: IBM Research / IBM Granite

One server for three kinds of request

IBM’s new Granite Switch 4.1 configuration brings chat, decision scoring and English speech transcription into a single hosted endpoint. The 8B and 30B checkpoints pair a Granite 4.1 base with a general-purpose Decision adapter, a Guardian safety adapter and the 470-million-parameter Granite Speech 5.0 TurboCTC encoder. Their model cards date the release to October 14; IBM’s explanatory announcement followed on October 15.

The practical pitch is fewer separately served components for an application that needs to hear a request, choose among options, check content and produce a response. An application selects an adapter through the chat template. Decision returns probabilities over labelled choices, while Guardian evaluates specified safety criteria. These are documented operating modes, not proof that the model will choose correctly in every domain.

Shared context is the efficiency argument

IBM says its activated LoRA adapters can reuse the base model’s cached context, avoiding repeated processing when a workflow switches from chat to a decision. Its announcement also claims approximately base-model-level GPU memory use and faster performance in a five-step retrieval workflow comparison. Those are IBM’s claims; AITrending has not reproduced the serving or benchmark results.

The cards list Transformers and vLLM backends, with Python 3.9 or newer, PyTorch 2.11 or newer, Transformers 5.16 through 5.17.0, and optional vLLM 0.26 through versions below 0.28. They declare Apache 2.0 licensing. Running a local endpoint still requires suitable hardware and the project’s inference dependencies; this is not an automatically provisioned free cloud API.

Speech is a transcription cascade

The audio encoder transcribes incoming speech and inserts that text into the chat context. It is English-only, produces lowercase transcripts without punctuation, and is not jointly trained with the language model. A transcription error can therefore change the request the model actually sees. This configuration should not be mistaken for native end-to-end audio understanding or speech generation.

The supplied Guardian adapter also does not include the library’s factuality detection, factuality correction or policy-guardrail adapters. Those require a different composition. IBM plans an Adapter SDK for training custom adapters in Mellea, while the current Decision adapter has no dedicated Mellea intrinsic yet. Those planned pieces remain separate from the downloadable checkpoints.

Related reading