IBM releases Granite Switch 4.1, putting chat, decisions and speech in one 8B or 30B model
IBM released Granite Switch 4.1 in 8B and 30B sizes on 14 October 2026, bundling a base model, two aLoRA adapters and a speech encoder under Apache 2.0.
By Emre AlemdarPublished
Tech AI LatestIBM released Granite Switch 4.1 on 14 October 2026 in 8B and 30B sizes under Apache 2.0. Each checkpoint combines a Granite 4.1 base model, a Decision adapter, a Guardian safety adapter and a 470M-parameter English speech encoder, so one endpoint can handle chat, classification decisions and transcription. IBM says the adapters share the base model's KV cache and use about the same GPU memory as the base model alone. Both sizes have a 131,072-token context.
Key facts
- Released: 14 October 2026 (model card); IBM launch post on 15 October
- Sizes: 8B and 30B base models, plus the 470M Granite Speech 5.0 TurboCTC encoder
- License: Apache 2.0; context window 131,072 tokens
- Adapters: Decision aLoRA and Guardian Core aLoRA, switched on by control tokens
- Decision Index v0.3 fast subset: 42.9 overall (8B) and 50.1 overall (30B)
- Runs in Hugging Face Transformers and in vLLM through a plug-in
What IBM released
IBM released Granite Switch 4.1 in 8B and 30B sizes. The model card for the 8B version gives a release date of 14 October 2026 and an Apache 2.0 license. Each checkpoint bundles a Granite 4.1 base model, a Decision adapter, a Guardian Core adapter and the 470M-parameter Granite Speech 5.0 TurboCTC encoder. The context window is 131,072 tokens. The base model covers 12 languages, including English, German, French, Japanese, Arabic and Chinese, while speech-to-text is English only.
IBM researchers Kate Soule, Subhajit Chaudhury, Hendrik Strobelt and Luis A. Lastras described the release in a launch post on 15 October. The 8B card describes the approach as coarse-grained expert switching: one checkpoint and one endpoint for chat, decisions and speech.
How the switching works
The adapters are activated LoRAs, or aLoRAs, and they turn on through control tokens in the chat template. Because they reuse the KV cache from the base model's generation, IBM says a later decision costs little extra compute, and the composed model takes about the same GPU memory as the base model by itself. The Decision adapter returns a probability distribution over labelled options in one forward pass. The Guardian Core adapter checks for harm, jailbreaking, profanity, violence, sexual content, social bias and unethical behavior.
The speech encoder is spliced into the chat context in place of audio input. The card notes that it was not trained jointly with the language model. The models run in Hugging Face Transformers and in vLLM through a plug-in from the Granite Switch repository on GitHub, and the launch post says llama.cpp also has native support. IBM first described Project Granite Switch as an experimental toolkit in a June 2026 IBM Research post; the October release ships ready-made checkpoints.
What IBM claims about performance
On the public fast subset of Decision Index v0.3, with 1,068 rows, the 8B model scores 42.9 overall and 59.8 on tools, and the 30B model scores 50.1 overall and 76.5 on tools. IBM says the Guardian adapter beats several named decision models on average on safety decision tasks, but the launch post gives no figures for that comparison. It also says a custom 3B Switch model playing Doom made 60 to 100 decisions per second. These are IBM's own results.
What to watch next
The launch post points to Granite Libraries, a repository of adapters that can be composed into custom Switch models, so third-party combinations are the next thing to watch. Hosted pricing for Granite Switch is not stated in the announcement, and independent safety benchmark results with numbers are not yet published. The release is listed on the AI model release tracker. An earlier open-weight release this month was Aleph Alpha's Kolibri-1, and more stories are under Open Source AI.
What changed
| Before | Granite 4.1 base models and Granite Libraries adapters shipped separately, and Project Granite Switch was an experimental composition toolkit (June 2026) |
|---|---|
| After | Granite Switch 4.1 ships ready-made 8B and 30B checkpoints with Decision and Guardian adapters and an English speech encoder, under Apache 2.0 |
| Who it affects | Developers running chat plus classification or safety screening on one GPU endpoint; teams adding English transcription to Granite deployments |
Questions people ask
- What is IBM Granite Switch 4.1?
- A model release in 8B and 30B sizes that bundles a Granite 4.1 base model, Decision and Guardian adapters and a speech encoder in one checkpoint, so one endpoint serves chat, decisions and English speech-to-text.
- What license does Granite Switch 4.1 use?
- Apache 2.0, according to the model cards on Hugging Face.
- Does Granite Switch 4.1 need more GPU memory than Granite 4.1?
- IBM says the composed model uses approximately the same GPU memory as the base model alone, because the adapters are low-rank and share the base model's KV cache.
Sources
- granite-switch-4.1-8b model card (IBM Granite on Hugging Face, 14 Oct 2026)
- Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint (IBM Granite (Hugging Face blog), 15 Oct 2026)
- granite-switch-4.1-30b model card (IBM Granite on Hugging Face, 14 Oct 2026)
- How to build AI more like software (IBM Research, 4 Jun 2026)
Reference pages
- Granite Switch 4.1 model
- IBM company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.
