Granite Switch 4.1
IBM · Released 15 Oct 2026
Modular Granite Switch models (8B and 30B) that combine a general-purpose Decision aLoRA, a specialized Guardian safety-decision adapter, and speech-to-text in one hosted endpoint, so a single model can move between chat, decisions, and transcription.
- Maker
- IBM
- Release date
- 15 Oct 2026
- First seen
- 15 Oct 2026
- Access
- Model cards and weights via IBM's Granite Switch resources on Hugging Face; runs in vLLM with a plug-in and has native inference support in llama.cpp. IBM also published a Speech-to-Decisions demo space.
- Pricing (as publicly stated)
- Released as compose-it-yourself open models; no public hosted list price was stated in the 15 Oct 2026 announcement.
- Modalities
- text, speech
- Question types
- general-purpose decisions (v1/systemone-compatible), specialized safety decisions
- Latency (as stated)
- IBM said a custom 3B Granite Switch model used in a Doom demo made 60-100 decisions per second, compared with Jev's reported 10 decisions per second, and that it can also narrate and answer questions while playing.
- Official announcement
- https://huggingface.co/blog/ibm-granite/granite-switch-4-1
Benchmark claims
These are claims as published by the maker or in the linked official post. We have not independently rerun every benchmark.
IBM said the specialized Guardian aLoRA, on average, outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks. It also said a Granite-Switch-4.1-30B setup with four specialized aLoRAs completed a 5-step RAG workflow much faster than Jev-27B in an internal comparison.
Names and marks belong to their owners. This page is an independent summary, not a product page from IBM.