AI PCs · November 20, 2026 · AMD
AMD and Unsloth Expand Local NPU Inference Access
Reported company announcement
AMD’s November 19 announcement says more than 100 compatible agentic AI applications and workflows can access local model inference through Unsloth Desktop on supported Ryzen AI PCs. AMD XDNA 2 neural processing units are the first NPUs supported by that application.
Applications send model requests through Unsloth Desktop; Lemonade and FastFlowLM handle serving and inference on the NPU. Other application functions still use CPU and GPU resources as needed.
The accompanying AMD technical guide describes the Windows installation path and notes that memory requirements vary by model. Its full-context Qwen 3.6 35B A3B example requires about 28 GB available to the NPU, implying 64 GB of system memory under the stated default shared allocation.
Analysis
For AMD watchers, the development broadens software access to existing dedicated AI hardware. The “100+” figure describes compatible applications and workflows; it is not a measured count of customers or new chips sold.
Local inference also does not establish that every connected application operates entirely offline. Hardware, model fit and the rest of the workflow still matter. We infer no sales, battery-life improvement for every workload, or stock-price direction and make no buy or sell recommendation.
Related reading
Browse the market-report index or return to the clock dashboard.