AMD Unveils SD 3 Medium: The World's First BF16 Stable Diffusion Model Optimized for XDNA 2 AI NPUs, Now Available via Amuse 3.1

AMD has officially launched the world's first BF16 (block FP16) Stable Diffusion model tailored specifically for its latest-generation XDNA 2 NPUs. Named SD 3 Medium, this new model is engineered to deliver faster, high-quality, local text-to-image generation with reduced memory footprint, setting a new standard for on-device generative AI performance. Read the official announcement.

Following the company’s success with SDXL Turbo, first introduced at Computex 2024, which blended FP16 accuracy with INT8 performance through a collaboration with Stability AI available on Hugging Face, AMD now introduces SD 3.0 Medium — a Block FP16 model offering significantly enhanced image quality. This new model is fully optimized for AMD Ryzen AI laptops powered by the XDNA 2 NPU architecture, and is already available to try using the Amuse 3.1 Beta app by Tensorstack. Try SD 3 Medium here.

To use SD 3 Medium on your local device, you need an AMD Ryzen AI 300 series or Ryzen AI MAX+ laptop equipped with a minimum of 50 TOPs NPU performance and at least 24GB of system RAM. The setup is straightforward:

  1. Download and install the latest Adrenalin Driver from AMD.

  2. Download and install Amuse 3.1 Beta.

  3. Launch in EZ Mode, set the slider to HQ, and enable “XDNA 2 Stable Diffusion Offload”.

One of the key advantages of SD 3 Medium is its minimal resource consumption — it requires just 9GB of system memory, which allows it to run efficiently even on laptops under 32GB. At the same time, it preserves the fidelity and accuracy of traditional FP16 image generation. In addition, SD 3 Medium employs a 2-stage NPU-powered pipeline to enhance the base image from 1024x1024 pixels (2MP) to a stunning 2048x2048 pixels (4MP), delivering print-quality images directly from your device.

This breakthrough empowers creators, marketers, and designers to generate high-resolution visuals locally, with no need for cloud connectivity, subscription fees, or third-party rendering services. Whether you're working in the office, traveling, or completely offline, AMD’s XDNA 2 NPU ensures seamless generative AI workloads with unmatched power efficiency and performance.

For AMD, SD 3 Medium is not just a technical achievement — it’s a milestone in redefining the AI PC. It reflects AMD’s vision of enabling high-quality, custom-tailored AI image generation at the edge, giving professionals the flexibility and control to iterate creative designs quickly and securely from any location. Whether you’re plugged in at your desk or flying at 30,000 feet, Ryzen AI + SD 3 Medium delivers the performance you need without compromise.


What are your thoughts on running full Stable Diffusion models locally without relying on the cloud? Would you use SD 3 Medium for your creative workflows? Share your opinion in the comments below.

Angel Morales

Founder and lead writer at Duck-IT Tech News, and dedicated to delivering the latest news, reviews, and insights in the world of technology, gaming, and AI. With experience in the tech and business sectors, combining a deep passion for technology with a talent for clear and engaging writing

Next
Next

NVIDIA Accelerates AI Revolution: Blackwell Ultra Production Begins While Rubin-Based Vera Rubin Servers Near Final Design Phase