Skip to content

How-ToCPU

How to Run Local AI with Lemonade on AMD Ryzen AI and Radeon PCs

How to install Lemonade, the free open-source local AI server with AMD optimisations, on Windows 11; run a first model with lemonade run; switch between Vulkan and ROCm on Radeon GPUs; use Hybrid or FastFlowLM models on Ryzen AI 300/400 NPUs; and connect apps to its OpenAI-compatible API on port 13305.

By , Editorial team

4 min read

How to Run Local AI with Lemonade on AMD Ryzen AI and Radeon PCs

Most local AI apps are built first for NVIDIA graphics cards. Lemonade is different: it is a free, open-source local AI server whose developers describe it as "built by the community for every PC, with optimizations by AMD engineers to get the most from Ryzen AI, Radeon, and Strix Halo PCs". It can run chatbots on a Radeon graphics card, on the integrated graphics of a Ryzen laptop or mini PC, and on the NPU in Ryzen AI 300 and 400-series chips, and any app that speaks the OpenAI API can use it.

This guide follows Lemonade's documentation and AMD's official getting-started playbook, checked on 14 November 2026. Forge PC has not benchmarked Lemonade, so we do not quote speeds.

What you need

  • Windows 11 or a supported Linux distribution (AMD lists Ubuntu 24.04 or later, Fedora and Debian).

  • 16GB of RAM for the small model in AMD's guide (about 3GB), and 32GB or more for larger models such as the roughly 20GB coding model AMD uses later in its playbook. Our RAM for local AI guide explains why.

  • About 4GB to 30GB of free disk space, depending on the models you download.

  • The latest AMD graphics driver. In AMD Software: Adrenalin Edition, go to Driver and Software > Manage Updates. Our driver update guide has more detail.

  • Optional: an XDNA 2 NPU. AMD lists Ryzen AI 300, 400 and Max 300-series processors and the Z2 Extreme. You also need the NPU driver from AMD's Ryzen AI Software instructions.

Not on AMD? Lemonade's FAQ says it still works: Intel and NVIDIA systems get CPU inference, with GPU support through the llama.cpp Vulkan backend, and Lemonade's GitHub page lists a CUDA backend for NVIDIA GPUs from the Turing generation onwards.

Step 1: Install Lemonade

  1. Download the installer from lemonade-server.ai. AMD says two Windows installers are published: lemonade.msi, which includes the server and the desktop app (use this one), and lemonade-server-minimal.msi, which is the server only.

  2. Run the .msi file. The lemonade command is added to your PATH, and the server is expected to run in the background automatically.

  3. Open a terminal and type lemonade --version. If you see a version number, it is installed. lemonade status confirms the server is running.

Step 2: Run your first model

In a terminal, run:

lemonade run Gemma-4-E2B-it-GGUF

AMD's playbook says this downloads the model (about 3GB) from Hugging Face if needed, starts the Lemonade server on port 13305 and opens the Lemonade app so you can chat. If you installed the minimal package, open http://localhost:13305 in your browser instead. Use lemonade list to see available models and lemonade pull followed by a model name to download one without starting it.

Step 3: Choose Vulkan or ROCm on a Radeon GPU

Lemonade runs GGUF models through llama.cpp. On AMD graphics you can choose the Vulkan backend or ROCm (on GPUs that ROCm supports):

  • lemonade run <model> --llamacpp vulkan starts a model with Vulkan.

  • lemonade run <model> --llamacpp rocm starts it with ROCm.

  • lemonade config set llamacpp.backend=rocm makes ROCm the default.

If one backend gives you errors, try the other. Our Radeon local AI guide explains the difference between ROCm and Vulkan in more detail.

Step 4: Use the NPU on a Ryzen AI PC

On a Ryzen AI 300 or 400-series PC, AMD describes three NPU modes, all behind the same API:

Mode

How it works

Example from AMD

Hybrid (NPU + iGPU)

The NPU processes your prompt, the integrated GPU generates the reply

Qwen3-4B-Hybrid

NPU-only

The whole model runs on the NPU

Qwen-2.5-7B-Instruct-NPU

FastFlowLM (FLM)

An engine optimised for XDNA 2 NPUs

qwen3.5-4b-FLM

AMD says running on the NPU "is more power-efficient than using the GPU", which suits background tasks and laptops on battery. To try hybrid mode, run lemonade run Qwen3-4B-Hybrid. For FLM models on Windows, open Backends Manager in the app and install the FastFlowLM NPU backend. Lemonade's FAQ notes that Ryzen AI Software's NPU and hybrid modes currently work only on Windows, and that a hybrid model shows little NPU activity with short prompts, because the NPU only handles the prompt-processing stage. For background on NPUs, see what an NPU is.

Step 5: Connect other apps

Lemonade exposes an OpenAI-compatible API at http://localhost:13305/api/v1. Any app or script that lets you set a custom OpenAI base URL can use it, which is the same idea as the local servers in our Ollama API and LM Studio server guides. Lemonade's GitHub page says it also offers Anthropic- and Ollama-compatible APIs.

Our verdict

Lemonade is the most AMD-friendly way to run local AI: install lemonade.msi from lemonade-server.ai on Windows 11, run lemonade run Gemma-4-E2B-it-GGUF for a first chat, switch between Vulkan and ROCm with the --llamacpp flag on a Radeon GPU, and try Hybrid or FastFlowLM models on a Ryzen AI 300 or 400-series NPU. Other apps can connect at http://localhost:13305/api/v1. AMD recommends 16GB of RAM for small models and 32GB or more for larger ones. Forge PC has not benchmarked Lemonade; this guide follows Lemonade's and AMD's official documentation.

The bottom line

Lemonade is the most AMD-friendly way to run local AI: install lemonade.msi from lemonade-server.ai on Windows 11, run `lemonade run Gemma-4-E2B-it-GGUF` for a first chat, switch between Vulkan and ROCm with the `--llamacpp` flag on a Radeon GPU, and try Hybrid or FastFlowLM models on a Ryzen AI 300 or 400-series NPU. Other apps can connect at `http://localhost:13305/api/v1`. AMD recommends 16GB of RAM for small models and 32GB or more for larger ones. Forge PC has not benchmarked Lemonade; this guide follows Lemonade's and AMD's official documentation.

Sources

  1. 01GitHub, lemonade-sdk/lemonade repository
  2. 02AMD, Getting Started with Lemonade (AMD AI Playbooks)
  3. 03Lemonade, Frequently Asked Questions

About the author

Forge PC Staff

Editorial team

The shared byline for articles researched, written and edited by the Forge PC editorial team. How we source, check and correct our work is set out in our editorial policy.

More from Forge PC Staff

Read next

HDMI 2.1 vs HDMI 2.2: What Changes for PC Monitors, Graphics Cards and Cables

Versus

HDMI 2.1 vs HDMI 2.2: What Changes for PC Monitors, Graphics Cards and Cables

HDMI 2.2 doubles HDMI's maximum bandwidth from 48Gbps to 96Gbps, enabling uncompressed 4K 240Hz and 8K 60Hz with full chroma. Its Ultra96 label can mean 64, 80 or 96Gbps, so check the real figure, and use a certified Ultra96 HDMI Cable for full bandwidth. For most PCs today, HDMI 2.1 or DisplayPort is still enough.

Forge PC Staff4 min

How to Set the Context Length in Ollama (and Why It Matters for VRAM)

How-To

How to Set the Context Length in Ollama (and Why It Matters for VRAM)

Ollama's context length is how many tokens a model can see at once. Its default depends on VRAM (4k under 24GiB), and Ollama suggests at least 64,000 for agents, coding and web search. Change it with the app's slider, OLLAMA_CONTEXT_LENGTH, /set parameter num_ctx or the API, then check ollama ps shows 100% GPU.

Forge PC Staff4 min

Windows Will Let You Link PCs Over USB4 to Run Bigger Local AI Models

News

Windows Will Let You Link PCs Over USB4 to Run Bigger Local AI Models

Microsoft has announced a developer preview of RDMA over USB4 for Windows, letting two or more USB4 PCs split one AI model across their combined memory and GPUs with no special network hardware. Partner figures include about 18.5 tokens per second for a 235B model on three Ryzen AI Max mini PCs. It needs test signing and Secure Boot off.

Forge PC Staff3 min

Unsloth Desktop Adds AMD Ryzen AI NPU Support for Local AI Agents

News

Unsloth Desktop Adds AMD Ryzen AI NPU Support for Local AI Agents

AMD and Unsloth have added NPU support to Unsloth Desktop, so Ryzen AI PCs with an XDNA 2 NPU can serve local models to more than 100 compatible agent apps through an OpenAI-compatible API, using Lemonade and FastFlowLM. It works on Windows and, per Phoronix, Linux. AMD's 35B example model needs a 64GB system.

Forge PC Staff3 min