⚡ FastFlowLM (FLM) — Unlock Ryzen™ AI NPUs
Run large language models — now with Vision, Audio, Embedding and MoE support — on AMD Ryzen™ AI NPUs in minutes.
No GPU required. Faster and over 10× more power-efficient. Supports context lengths up to 256k tokens. Ultra-Lightweight (17 MB). Installs within 20 seconds.
📦 The only out-of-box, NPU-first runtime built exclusively for Ryzen™ AI.
🤝 A familiar single-command CLI — deeply optimized for NPUs.
✨ From Idle Silicon to Instant Power — FastFlowLM Makes Ryzen™ AI Shine.
FastFlowLM (FLM) supports all Ryzen™ AI Series chips with XDNA2 NPUs (Strix, Strix Halo, Kraken, and Gorgon Point).
🔗 Quick Links
🔽 Download | 📊 Benchmarks | 📦 Model List
🐧 Linux Getting Started Guide
📖 Docs | 📺 Demos | 💬 Discord
🚀 Quick Start
A packaged FLM Windows installer is available here: flm-setup.msi. For more details, see the release notes.
📺 Watch the quick start video (Windows)
[!IMPORTANT]
⚠️ Use the latest AMD NPU driver — 32.0.203.311 or above (check via Task Manager→Performance→NPU or Device Manager). Earlier versions are no longer supported.
⚙️ Tip:
After installation, open PowerShell (Win + X → I). To run a model in terminal (CLI Mode):
flm run llama3.2:1b
Notes:
- Internet access to HuggingFace is required to download the optimized model kernels.
- Sometimes downloads from HuggingFace may get corrupted. If this happens, run
flm pull <model_tag> --force (e.g. flm pull llama3.2:1b --force) to re-download and fix them.
- By default, models are stored in:
- Windows:
C:\Users\<USER>\.flm\models\
- Linux:
~/.config/flm/
- During installation on Windows, you can select a different base folder (e.g., if you choose
C:\Users\<USER>\flm, models will be saved under C:\Users\<USER>\flm\models\).
- On Linux, you can override the default location by setting the
FLM_MODEL_PATH environment variable.
- To disable the startup version check, set
FLM_DISABLE_UPDATE_CHECK=1.
- ⚠️ If HuggingFace is not accessible in your region, manually download the model (check this issue) and place it in the chosen directory.
🎉🚀 FastFlowLM (FLM) is ready — your NPU is unlocked and you can start chatting with models right away!
Open Task Manager (Ctrl + Shift + Esc). Go to the Performance tab → click NPU to monitor usage.
⚡ Quick Tips:
- Use
/verbose during a session to turn on performance reporting (toggle off with /verbose again).
- Type
/bye to exit a conversation.
- Run
flm list in PowerShell to show all available models.
To start the local server (Server Mode):
flm serve llama3.2:1b
The model tag (e.g., llama3.2:1b) sets the initial model, which is optional. If another model is requested, FastFlowLM will automatically switch to it. The local server runs on port 52625 (default).
📰 In the News
08/11/2026 🎉 FLM is now part of ROCm (v1.0.0) — the repo has moved to AMD's open-source ROCm organization.
08/11/2026 🎉 FLM releases its first SmolVLA model (v1.0.0) — a Vision-Language-Action robotics policy running on the NPU. See the model card and benchmarks.
07/17/2026 🎉 FLM is now part of AMD news. Read our story — from a 2025 university project to AMD.
03/11/2026 🎉 FLM now supports Linux 🐧 ! To get started, check out the quick start guide or the Lemonade Server docs, and watch the short video for a quick walkthrough of FLM on Linux via Lemonade 🍋.
10/01/2025 🎉 FLM was integrated into AMD's Lemonade Server 🍋. Watch this short demo about using FLM in Lemonade.
🧠 Local AI on NPU
FLM makes it easy to run cutting-edge LLMs (and now VLMs) locally with:
- ⚡ Fast and low power
- 🧰 Simple CLI and API (REST and OpenAI API)
- 🔐 Fully private and offline
No model rewrites, no tuning — it just works.
✅ Highlights
- Runs fully on AMD Ryzen™ AI NPU — no GPU or CPU load
- Lightweight runtime (17 MB) — installs within 20 seconds, easy to integrate
- Developer-first flow — a familiar single-command CLI, optimized for NPU
- Support for long context windows — up to 256k tokens (e.g., Qwen3-4B-Thinking-2507)
- No low-level tuning required — You focus on your app, we handle the rest
📄 License
💬 Have feedback/issues or want early access to our new releases? Open an issue or Join our Discord community
🙏 Acknowledgements
🛠️ Building from Source
For developers who want to build FastFlowLM from source, we provide CMake presets for a convenient and consistent build experience.
To build the NPU engines, kernels and weight converter themselves, see flm-dev-kit/README.md.
Prerequisites
- Git
- CMake (version 3.22 or higher)
- A C++20 compatible compiler (e.g., GCC, Clang, MSVC)
- Ninja (recommended)
Build Instructions
More details on the exact procedure, with dependencies to be installed, for Linux can be found in linux-getting-started.md.
Clone the repository:
git clone --recursive https://github.com/ROCm/FastFlowLM.git
cd FastFlowLM/src
Configure CMake using presets:
For Linux:
cmake --preset linux-default
This will configure the build to install to /opt/fastflowlm.
To build the Gemma 4 engine from source instead of using the prebuilt engine:
cmake --preset linux-default -DFLM_BUILD_GEMMA4E=ON
cmake --build build
CMake caches this option. To switch back to the prebuilt engine, reconfigure with -DFLM_BUILD_GEMMA4E=OFF or use a fresh build directory.
For Windows (in a developer command prompt):
cmake --preset windows-default
Add -DFLM_BUILD_GEMMA4E=ON to build the Gemma 4 engine from source.
Build the project:
cmake --build build
Install the project (optional):