Deploy large language models (LLMs) on AMD Ryzen™ AI NPUs—in minutes.

Think Ollama, but purpose-built for the great AMD Ryzen™ NPUs.

​

FastFlowLM project runs LLaMA, DeepSeek-R1, and more—hardware-optimized for lightning-fast, low-power, private and always-on AI on the silicon already in your AI PC.

​​​

👨‍💻 Made for Developers Building Local AI Agents

• 🧠 No Low-Level Tuning Needed

  • You don’t need to understand NPU internals—just run your model, we handle the low-level hardware optimization.

• 🧰 Works Like Ollama—But For AMD Ryzen™ NPU AI

  • CLI and API simplicity developers love, with the efficiency and control of native NPU execution.

• 💻 No GPU or CPU Burden

  • FastFlowLM runs entirely on the Ryzen™ NPU, freeing system resources for other applications.

• 📏 Full Context Support

  • ​All FastFlowLM models support full context windows—up to 128k tokens on LLaMA 3.1/3.2—so you can run long-form reasoning and RAG without compromise.​

• ⚡ Performance That Speaks for Itself
  * ​Compared to AMD Ryzen™ AI Software 1.4 (GAIA or Lemonade):

  • LLM Decoding Speed (TPS: Tokens per Second)
    🚀 Up to 14.2× faster in LLM decoding (vs NPU-only baseline)
    🚀 Up to 16.2× faster in LLM decoding (vs hybrid iGPU+NPU baseline)

  • Power Efficiency
    🔋 Up to 2.66× more power efficient in LLM decoding (vs NPU-only)
    🔋 Up to 11.38× more power efficient in LLM decoding (vs hybrid)
    🔋 Up to 3.4× more power efficient in LLM prefill (vs NPU-only or hybrid mode)

  • Latency (LLM Prefill Speed)
    🚀 Matches or exceeds the Time to First Token (TTFT) of the NPU-only or hybrid mode