The Pulse
IBM Puts Chat, Decisions and Speech in Granite Switch 4.1
IBM’s Granite Switch 4.1 combines chat, decision and safety adapters with an embedded speech-recognition encoder in 8B and 30B checkpoints. The speech path transcribes English audio into text for the language model, with lowercase, unpunctu
AI.info Team · Written with AI ·
One IBM model endpoint can take a spoken question, transcribe it, respond in chat and switch to a decision or safety-check task. IBM published details of Granite Switch 4.1 on October 15, following model-card release dates of October 14 for its 8B and 30B versions. The checkpoints combine a Granite 4.1 language model with decision and safety adapters and an embedded speech-recognition encoder.
Two adapters, one speech encoder
The 8B and 30B versions use a decision adapter for choosing among labeled options and a Guardian adapter for detecting safety risks. IBM says the decision adapter returns a probability distribution over the supplied options, while Guardian can check for risks including harmful content, jailbreaking, violence and social bias. The speech component is Granite Speech 5.0 TurboCTC, a 470-million-parameter encoder for English transcription.
The transcription step feeds text into the language model’s chat context; the model card says the speech encoder and language model have not been jointly trained. That distinction makes the feature a speech-to-text path into a text model, not a single model trained to understand audio and language together. The card also flags English-only input and lowercase, unpunctuated transcription output.
“Check out our Speech-to-Decisions HF space where granite-switch-4.1-8b transcribes audio, generates summaries, screens for harmful content, and tests 15 other user-defined flags—all running on a single endpoint in real time.”
IBM Granite authors Kate Soule, Subhajit Chaudhury, Hendrik Strobelt and Luis A. Lastras write
IBM says the bundle fits base-model memory
IBM says the adapters and speech technology occupy approximately the same GPU VRAM as the underlying base model by itself. The company attributes that to adapters that can reuse the base model’s key-value cache across requests, alongside a co-located speech encoder. The design aims to avoid running separate model copies for chat, decisions and transcription, though the release does not publish a GPU-memory figure or a direct memory comparison.
Both checkpoints are listed under the Apache 2.0 license and can be run with Hugging Face Transformers or vLLM. They are composed from Granite 4.1 base models with 8 billion and 30 billion parameters, respectively. Developers activate the adapters through control tokens in the chat template, and IBM says the 8B checkpoint can also summarize transcripts and check more than 15 generic flags.
Decision claims need benchmark detail
IBM says its Guardian adapter averages better than Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety-related decision tasks. The announcement does not provide score tables or the evaluation details needed to assess that comparison, so it is a company-reported result rather than a fully documented benchmark in the post. IBM also describes a separate five-step retrieval workflow race against Jev-27B, but gives no elapsed-time figures there.
The release’s practical boundary is clearer: Granite Switch puts text generation, labeled-option decisions, safety checks and English transcription behind one model interface. Speech input still passes through a separate embedded encoder, and its output arrives as lowercase text without punctuation.