Skip to content

How-ToCPU

How to Run a Local LLM on an AMD Ryzen AI NPU With FastFlowLM

FastFlowLM runs language models on the XDNA 2 NPU in Ryzen AI chips. Update the NPU driver to 32.0.203.311 or later, install flm-setup.msi, test with flm run llama3.2:1b and use flm serve for an OpenAI-compatible local API.

By , Systems & Technical Editor

3 min read

How to Run a Local LLM on an AMD Ryzen AI NPU With FastFlowLM

Most local AI tools run language models on the graphics card or CPU. FastFlowLM (FLM) instead targets the NPU in AMD's Ryzen AI chips, which its developers say is over 10 times more power-efficient; AMD adds that offloading to the NPU can leave CPU and GPU resources free for other apps. The README describes it as ultra-lightweight (17MB) with a single-command CLI, and it can also run an OpenAI-compatible local server for other apps. This guide follows FastFlowLM's GitHub instructions and AMD's November 2026 blog posts.

What you need

FastFlowLM's README says it supports all Ryzen AI chips with XDNA 2 NPUs: the Strix, Strix Halo, Kraken and Gorgon Point families. The README describes FLM as built exclusively for Ryzen AI, so Intel and Qualcomm NPUs aren't covered; check that your chip has an XDNA 2 NPU. If you're not sure what's in your PC, our explainer on what an NPU is and our guide to checking NPU usage in Windows 11 will help.

You also need enough memory. In its Unsloth Desktop post, AMD gives the example of Qwen 3.6 35B A3B at full context needing around 28GB available to the NPU; with Windows' default 50% shared allocation, it says that means 64GB of system RAM. Smaller models need far less.

Step 1: Update the NPU driver

The README says FLM needs AMD NPU driver version 32.0.203.311 or above, and that earlier versions are no longer supported. Check the version in Task Manager (Performance, then NPU) or Device Manager. FastFlowLM recommends trying Windows Update or AMD's driver download first.

Step 2: Install FastFlowLM

Download the packaged Windows installer, flm-setup.msi, from the FastFlowLM GitHub releases and run it. During installation you can choose a different base folder for models; by default they're stored in C:\Users\.flm\models. If you're short of space on C:, pick another drive here.

Step 3: Run your first model

Open PowerShell (Win + X, then I) and run a small model to test everything works:

flm run llama3.2:1b

The README says internet access to Hugging Face is required to download the optimised model kernels. If a download gets corrupted, the README says to re-download it with flm pull llama3.2:1b --force. AMD's post also uses flm list to show available models.

Step 4: Serve it to other apps

To use FLM from other tools, start its local server:

flm serve llama3.2:1b

AMD's FastFlow post says flm serve exposes the model through an OpenAI-compatible local API, so apps that let you set a custom OpenAI-style endpoint can use it. AMD's 19 November post also says Unsloth Desktop now supports AMD XDNA 2 NPUs through Lemonade and FastFlowLM; see our Unsloth Desktop NPU story.

Optional: the new GPU path

FastFlow v1.0.13, released on 13 November 2026, added an early-access GPU path using AMD's HIP EP backend. Models with an -oga-hipep tag, such as qwen3.5-oga-hipep:9b, run on the integrated GPU instead of the NPU. AMD says this path still requires a supported NPU in the PC and that features and performance may change. We covered the announcement in our FastFlowLM GPU support story.

Tips

Our verdict

If you own a Ryzen AI laptop or mini PC with an XDNA 2 NPU, FastFlowLM is the simplest way to put that NPU to work: update the driver to 32.0.203.311 or later, install flm-setup.msi, and run flm run to test or flm serve to share a model with other apps. On anything else, stick with GPU-based tools.

The bottom line

If you own a Ryzen AI laptop or mini PC with an XDNA 2 NPU, FastFlowLM is the simplest way to put that NPU to work: update the driver to 32.0.203.311 or later, install flm-setup.msi, and run `flm run` to test or `flm serve` to share a model with other apps. On anything else, stick with GPU-based tools.

Sources

  1. 01FastFlowLM, FastFlowLM GitHub repository and README
  2. 02AMD, FastFlow now runs on AMD GPUs (Nov 2026)
  3. 03AMD, AMD x Unsloth: Unlocking the World's First NPU for Unsloth Desktop (19 Nov 2026)

About the author

Marcus Thorne

Systems & Technical Editor

Marcus writes Forge PC's step-by-step build, troubleshooting and system optimisation guides. His background spans Windows kernel and power-state diagnostics, UEFI/BIOS configuration, NVMe storage arrays, ATX 3.1 power delivery and running local AI and LLM workloads on consumer graphics cards.

More from Marcus Thorne

Read next

How to Use Ollama With Obsidian for Private, Local AI Notes

How-To

How to Use Ollama With Obsidian for Private, Local AI Notes

Use the Copilot for Obsidian plugin with Ollama to chat with and search your notes locally. Pull a chat and embedding model, set OLLAMA_ORIGINS to app://obsidian.md*, add the model with the ollama provider and raise the context window.

Marcus Thorne3 min

How to Connect a PS5 DualSense Controller to Your PC

How-To

How to Connect a PS5 DualSense Controller to Your PC

A PS5 DualSense connects to a Windows PC by USB-C or Bluetooth. PlayStation says haptics and the headset jack need a cable on PC; to pair wirelessly, hold Create and PS until the light bar blinks, then add it in Windows.

Marcus Thorne3 min

How to Use Ollama With n8n for Local AI Workflows

How-To

How to Use Ollama With n8n for Local AI Workflows

Run n8n in Docker, make Ollama listen on 0.0.0.0, and set the n8n Ollama credential's Base URL to http://host.docker.internal:11434. A Chat Trigger, Basic LLM Chain and Ollama Chat Model then make a first local AI workflow.

Marcus Thorne4 min