New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.New12,480+ smartphones indexed — free sandbox tier on every plan.Edge-cached JSON responses, typically under 100ms.SDKs for C#, TypeScript and Python, plus an MCP server for AI agents.
Get Free API Key
AI Hardware & Local Inference•4 min read•2026-11-19

AMD and Unsloth Bring XDNA 2 NPU Acceleration to Over 100 Agentic AI Apps via Unsloth Desktop

AMD and Unsloth introduce local XDNA 2 NPU acceleration in Unsloth Desktop for Windows and Linux, powering 100+ agentic AI apps and 35B-parameter MoE models via Lemonade and FastFlowLM.

Tech News Bot
DeviceSpecs Intelligence
Original Source Reference
#AMD#Unsloth#XDNA 2#Ryzen AI#NPU#Local AI#Agentic AI#FastFlowLM

AMD and Unsloth have announced a collaboration bringing local neural processing unit (NPU) acceleration to Unsloth Desktop, making AMD XDNA 2 NPUs the first NPUs supported by the platform across both Windows and Linux. Powered by Lemonade and FastFlowLM, the integration enables more than 100 agentic AI applications and coding tools to offload local LLM inference to Ryzen AI NPUs.

World-First NPU Support in Unsloth Desktop for Windows and Linux

Starting with AMD Ryzen AI laptops and desktops equipped with XDNA 2 and newer NPUs, users can now enable dedicated NPU execution directly inside Unsloth Desktop model hub. Rather than moving an entire agent stack onto the NPU, the architecture routes supported model inference through AMD open-source Lemonade and FastFlowLM (FLM) serving layers while keeping workflow orchestration and UI tasks on the CPU and GPU, preserving battery life for unplugged, always-on agents.

Connecting 100+ Agentic Applications via an OpenAI-Compliant API

Through the local Unsloth Desktop API, compatible developer tools and autonomous agents can tap into NPU inference via a standard OpenAI-compliant endpoint. AMD highlighted out-of-the-box compatibility across more than 100 agentic applications and workflows, including Hermes Agent, OpenClaw, Open WebUI, Cline, OMP, Copilot, and Visual Studio Code—allowing coding assistants and background agents to run persistently without taxing discrete GPUs.

Running 35B-Parameter Mixture-of-Experts Models on Ryzen AI Silicon

To demonstrate the capability, AMD showcased real-time local inference of a 35-billion-parameter Qwen 3.6 Mixture-of-Experts model (Qwen 3.6 35B A3B) running on an XDNA 2 NPU via Unsloth Desktop and FastFlowLM. Because full-context execution of the 35B model requires roughly 28GB of NPU-accessible memory, AMD noted that Windows systems using the default 50 percent shared memory allocation require 64GB of system RAM for full-context workloads.