AMD's Ryzen AI Software 1.9 lets applications switch task-specific model adapters while a quantized base remains loaded, giving developers a way to serve different workloads from one model.
AMD announced the update Nov. 2. The release adds dynamic adapter switching for selected text models, including Qwen2.5-3B, Qwen3-4B and Qwen3.5-4B. Applications can select an adapter during inference, when a trained model generates a response.
Low-Rank Adaptation, or LoRA, adds learned parameters for a particular task while leaving a model's original weights unchanged. The AMD workflow quantizes a base model with AMD Quark using INT4 AWQ, then trains and saves each adapter separately against that quantized base.
The base model and adapters are compiled as distinct artifacts. An application can load the base once and choose an adapter at runtime, so it can change tasks without loading the base model again.
Version 1.9 also extends context length and expands the supported model list. AMD's release notes list the additional model support.
