IBM has released 8B and 30B Granite Switch configurations combining a chat model, Decision and Guardian adapters, and Granite Speech transcription. They share one endpoint rather than requiring entirely separate serving systems.
The aLoRA approach can reuse the base model’s KV cache for subsequent decisions. Precomposed models and a Speech-to-Decisions demo are available, and teams can compose their own adapters. A planned Adapter SDK should not be described as already shipped.
Memory savings, ASR rankings and Jev comparisons are IBM’s reported results under its configurations. A Guardian adapter does not guarantee safety of the complete agent. Check licenses for models, libraries and components separately.
Composing specialists around a shared base offers an alternative to making separate calls to multiple models.