One endpoint, four jobs

Most teams that need chat, classification, safety screening and transcription run them as separate services, often from separate vendors. Granite Switch 4.1 is IBM's attempt to fold that stack into a single checkpoint. The base Granite 4.1 model handles the conversation. A Decision adapter takes a state and a question with labelled options and returns probabilities across those labels in one forward pass. A Guardian adapter screens for harm, jailbreak attempts, bias and other risk categories. An embedded speech encoder transcribes incoming audio, and the transcript simply takes the audio's place in the chat context.

The mechanism is the point. The adapters are Activated LoRAs trained against a shared, normalised KV cache, so an adapter can reuse the work the base model already did instead of recomputing it. Each one stays active across a span of tokens, is triggered by a control token the chat template inserts, and was developed and tested on its own, with no joint training. IBM's Speech-to-Decisions demo shows the shape of the intended product: the 8B model transcribes audio, summarises it, screens it for harmful content and checks 15 further user-defined flags, in real time, on one endpoint.

The claim to read carefully is the memory one

IBM says all of this takes approximately the same GPU memory as the base model by itself. That is plausible in the narrow sense: LoRA adapters are small next to an 8B or 30B base, and the speech encoder is 470 million parameters. It is also a vendor claim about its own packaging, stated with an "approximately" attached, and it describes memory, not speed, accuracy or operational simplicity. Anyone composing their own Switch from the adapter catalogue, which IBM invites, will land on a different total. Treat the figure as a design goal IBM says this configuration meets, not as a property of any Granite Switch you build yourself.

IBM published its own scorecard, including the losses

The Decision Index numbers are not launch-day fireworks. The 30B adapter scores 50.1 overall on IBM's 1,068-row public subset, the 8B adapter 42.9, and the category split is uneven: Tools reaches 76.5 at 30B while Arts sits at 26.4 and Knowledge at 32.4. In IBM's comparison of its specialised adapters against generalist decision models, the Granite entries win some categories by wide margins and lose others, including both InfoBench variants. IBM published those losing rows in its own table.

That candour continues in the limitations, which are the most useful part of the release. The speech-to-text path works only in English — a limit of the embedded encoder, not of the base model, which the card lists as working in twelve languages — and it returns lowercase text with no punctuation and no way to steer the decoding. Because the speech encoder was not trained jointly with the language model, a transcription error does not raise a flag; it quietly becomes the text the model and the adapters then reason over. Guardian-core is explicitly not a substitute for testing your own application. And the decision probabilities are, in the cards' framing, a prior that informs a decision, not a certified classification for fields the adapter was not trained on. A buyer who reads only the announcement post will miss all four points. A buyer who reads the cards gets a realistic deployment brief.

The practical angle

This release is aimed at teams already self-hosting open models for pipelines that mix conversation with labelled decisions: routing, screening, extraction checks, spoken requests that end in a classification. For them, one Apache 2.0 endpoint replacing a chat server, a classifier and a speech service is a real reduction in moving parts, and the composer tooling means the adapter set can match the job rather than the vendor's bundle. The costs to price in are the pinned software stack — Python 3.9 or later, PyTorch 2.11 or later, Transformers held between 5.16 and 5.17, and vLLM between 0.26 and 0.28 if you serve with it — the English-only speech path, and decision quality that IBM's own scores put in the middle of the pack.

If you need multilingual transcription, certified classification, or a hosted service with someone else carrying the operational risk, this changes little. Switch is infrastructure for builders, and it reads that way.

Sources

  1. [1] IBM Granite (primary), "Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint", Hugging Face, published 15 October 2026, 20:34 UTC (composition, same-memory claim, KV-cache reuse, Speech-to-Decisions demo, RAG-workflow and Doom demonstrations, composer tooling)Read source
  2. [2] IBM Granite Switch 4.1 8B model card (primary), ibm-granite/granite-switch-4.1-8b, Hugging Face (release date 14 October 2026 — Apache 2.0, 131,072-token context, Decision Index table, the 128-token-id per-call cap on logprob_token_ids and IBM's chunking workaround, version requirements, and the stated limitations)Read source
  3. [3] IBM Granite Switch 4.1 30B model card (primary), ibm-granite/granite-switch-4.1-30b, Hugging Face (release date 14 October 2026 — the same primary documentation for the 30B checkpoint, including its Decision Index table and limitations)Read source
  4. [4] Unite.AI, "IBM Releases Modular Granite Switch 4.1 Models in 8B and 30B Sizes", 15 October 2026 (corroboration — configuration, switching mechanism, adapter comparison figures, availability)Read source