Granite Switch 4.1
IBM's 8B and 30B Granite models bundling decision, safety guardian and speech-to-text adapters in one endpoint with shared KV cache.
Overview
Granite Switch 4.1, published by IBM Granite on Hugging Face on October 15, 2026, comes in 8B and 30B sizes. Each model bundles a general-purpose Decision aLoRA, a Guardian aLoRA for safety decisions and Granite-Speech-5.0-Turbo-CTC for speech-to-text. Because the adapters are Activated LoRAs, the KV cache from a base-model generation is reused for later adapter-based decisions, and IBM says the combined model needs about the same GPU VRAM as the base model alone. Custom Granite Switch models can be assembled from adapters in the Granite Libraries; an Adapter SDK for training adapters in Mellea is planned but not yet released. Granite Switch 4.1.1, a backwards-compatible patch published on October 29, 2026, swaps in an updated speech model (granite-speech-5.0-470m-turboctc, 10 languages, native punctuation and capitalization) and an updated Decision adapter in granitelib-core r1.0 that IBM says scores higher on the Jev Decision Index test set; the base model is unchanged.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 8B and 30B sizes with Decision aLoRA, Guardian aLoRA and speech-to-text in one model
- Activated LoRAs reuse the base-model KV cache across adapter calls
- About the same GPU VRAM as the base model alone (IBM-stated)
- vLLM plug-in on GitHub and native llama.cpp inference support
- Build custom Switch models by mixing adapters from the Granite Libraries
- Speech-to-Decisions demo Space on Hugging Face
- 4.1.1 patch (2026-10-29): 10-language ASR with native punctuation and an updated Decision adapter, no migration needed
Use cases
- Voice agents that transcribe, classify intent and route in one pass
- Guardrail and safety checks on agent outputs without a second model deployment
- Fast multi-step RAG workflows that mix generation and classification decisions
Pricing
Open weights on Hugging Face; license not stated in the announcement (check model cards); no hosted pricing stated.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 15 Oct 2026
- Record updated
- 3 Nov 2026
- Record last checked
- 3 Nov 2026