IBM Granite Switch 4.1: Chat, Decisions, Guardian and Speech From One Open Endpoint

IBM has released Granite Switch 4.1, a pair of open models that put chat, structured decisions, safety screening and speech-to-text behind a single endpoint. The 8B and 30B versions combine a Granite 4.1 base model with two adapters and a speech encoder, so one deployment can switch between jobs that teams usually serve with several models. The release date on the Hugging Face model card is October 14, 2026, and the launch post is dated October 15.

This is a tooling release more than a "bigger model" release, and most of the interesting claims come from IBM itself. Below: what is inside, how the adapter trick works, how to run it, and which numbers to treat as vendor-reported.

TL;DR: what you get

What is Granite Switch, and what changed in 4.1?

Per IBM, the Granite Switch project began last spring as a modular way to build models: take a frozen base model, attach small adapters, and let control tokens in the chat template pick which one runs. Version 4.1 ships two ready-made "batteries included" configurations, and IBM says anyone can compose their own from the Granite Libraries adapters using the open-source granite-switch repository.

Three capabilities are bundled in each checkpoint:

• Chat: the base Granite 4.1 model answers in natural language as usual. Our coverage of Granite 4.2 describes the reasoning line that sits alongside it.
• Decisions: a new general-purpose Decision adapter that picks among labelled options.
• Safety: a Guardian adapter that detects harm, jailbreaking, profanity, violence, sexual content, social bias and unethical behavior, per the model card.
• Speech: an embedded ASR model so audio can be transcribed on the same endpoint.

How activated LoRA lets one endpoint do several jobs

Diagram-style image of fine-tuning adapters attached to a base model, representing Granite Switch aLoRA adapters

A normal LoRA adapter modifies a model from the first token, so switching adapters means recomputing the whole prompt. An activated LoRA (aLoRA) only changes the model after an activation token, which means the keys and values the base model already computed stay valid. IBM's blog says exactly this: "the KV cache created from a base model generation can be reused for any subsequent adapter-based decisions."

On the model card, the adapters are triggered via control tokens in the chat template. The Decision adapter is invoked by a ### DECISION turn, and Guardian by a <guardian> marker. With vLLM you select an adapter with one request field, chat_template_kwargs with an adapter_name. Because the shared prefix is cached, a chat reply followed by a safety check or a decision is cheap compared with calling a second model that must re-read the conversation.

IBM also says all three technologies plus the base are composed into one model that uses "approximately the same amount of VRAM on a GPU as the base model would by itself." That is IBM's claim; we have not measured it.

What the Decision adapter returns

Decision models are a small but busy category: they take a state, a question and a fixed list of options, then answer with a probability distribution rather than free text. The card says the adapter reads the logits of the single-token option labels and takes a softmax, so one forward pass yields calibrated scores that your code can threshold.

IBM reports results on the public Decision Index fast subset (1,068 rows, edition v0.3 scoring), measured through a hosted vLLM endpoint and locally, agreeing within about 0.6 points:

These are IBM's numbers. The card does not include a side-by-side table against rivals, so we cannot tell you from it alone how 42.9 ranks. The blog post does make comparative claims: that the Guardian adapter "on average, outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks," and that a 30B Granite Switch with four specialist adapters finished a five-step RAG workflow faster than Jev-27B in a "race." The post gives no table for either, and IBM says more data will follow "in the upcoming weeks."

For context on the competitors, see our posts on Jev 1.14 and what TypeSafe actually shipped and Liquid AI's d1-Ultra decision model. Decision models are becoming a contested product category, and each vendor uses its own favorite benchmark.

The Doom demo, and how to read it

IBM's most eye-catching comparison: a custom 3B Granite Switch model made "split second decisions to navigate the Doom game" at 60 to 100 decisions per second, versus Jev's reported 10 decisions per second, citing a TypeSafe post. IBM adds that its model can also narrate and answer questions while playing.

Treat this as a speed demo, not a quality benchmark. It compares IBM's own tuned 3B model against a number reported by a competitor, on a task of IBM's choosing. It does show what the architecture is good at: many tiny, cache-sharing decisions per second. It does not show that the general Decision adapter is better at your business logic.

Speech on the same endpoint, with caveats

Audio is transcribed by Granite Speech 5.0 Turbo CTC, a 470M-parameter non-autoregressive conformer CTC encoder that IBM calls "the fastest ASR model on the OpenASR Leaderboard." The transcript is spliced into the chat context where the audio part was, so you send audio through the usual OpenAI-compatible chat call with an input_audio part.

The card is candid about limits: the speech encoder and the LLM are "not yet jointly trained," there is no trained projection coupling them, and output is English-only, lowercase and unpunctuated with no decode-time steering. So this is a cascade (ASR then LLM) wrapped in one API, not a native multimodal model. If you need punctuation or other languages, plan for a separate step. For other small on-device speech work, see Tether's 250M transcription model.

How to try it

1. Install the project: clone generative-computing/granite-switch and pip install -e ".[vllm]" (or ".[hf]" for Transformers, ".[audio]" for audio dependencies).
2. Serve it: python -m vllm.entrypoints.openai.api_server --model ibm-granite/granite-switch-4.1-8b --port 8000.
3. Call it with any OpenAI client pointed at http://localhost:8000/v1. Add the adapter name through extra_body to run Guardian or Decision.
4. For application code, IBM recommends Mellea, which wraps adapters as operations; the card's example is guardian.guardian_check(context, backend, criteria="harm", target_role="user"), returning a score where 0.5 or more means risk detected. Decision has no Mellea intrinsic yet, so use the lower-level path in the card.
5. To build your own, run the composer: python -m granite_switch.composer.compose_granite_switch with a base model, adapter libraries and --include-adapters.

The model had only 28 downloads on Hugging Face on release day, so there is essentially no independent testing yet.

What this means for what you build or pay

AI safety guardrails concept for the Granite Switch Guardian adapter screening prompts

• Fewer models to host. If your pipeline already runs chat, a moderation check and a router, one composed checkpoint removes two deployments and two sets of GPUs, at least on IBM's account.
• Cheap guardrails. A safety check that reuses the chat KV cache costs far less than re-sending the conversation to a separate classifier. Whether Guardian is accurate enough for your policy is something to test, not assume; it is a classifier, not a guarantee.- Apache 2.0. Commercial use is straightforward on the card's terms, which is not true of every "open" release; compare with the licensing discussion in our Granite 4.2 post.
• Extensibility. IBM says an Adapter SDK for training custom and third-party adapters in Mellea is planned, not shipped.

What is claimed and what is verified

Why it matters

The trend worth noticing is composition. Instead of one giant model that does everything, vendors are shipping a good-enough base plus swappable, cache-friendly specialists. That shifts the buying question from "which model is best" to "which adapters can I trust, and how do I test them." IBM bets that safety and decisions are the two adapters most teams will want first. It is a sensible bet, and the proof will come when third parties run Guardian against their own red-team sets and compare the Decision adapter with Jev and Liquid on their own tasks.

For other recent open releases to compare with, see NVIDIA's Nemotron 3.5 Super VL.

Details are as published by IBM on Hugging Face on October 14 and 15, 2026 and may change.

Related reading

• IBM Granite 4.2: open reasoning models with agentic RL
• Jev 1.14: what TypeSafe shipped
• Liquid AI d1-Ultra decision model
• Tether on-device speech recognition
• NVIDIA Nemotron 3.5 Super VL