Qwen3.8-Flash-Next 125B runs on AMD Strix Halo without a discrete GPU

🔥 Qwen3.8-Flash-Next 125B runs on AMD Strix Halo without a discrete GPU Qwen3.8-Flash-Next with 125 billion parameters was brought up on AMD Strix Halo — no discrete GPU needed. The hardware-software combo: 128 GB of unified memory, a HALOGEN layer, and Fedora 43. At each generation step, only 6 billion of the MoE parameters are active — that's what the whole setup hinges on. 🔗 Read on habr.com (https://habr.com/ru/articles/1098996/) 📖 Read the full story on ai-release.org (https://ai-release.org/posts/qwen3-8-flash-next-na-125b-zapustili-na-amd-strix-halo-bez-diskretnoy-1a2a64f3.html?utm_source=tg&utm_medium=channel&utm_campaign=hourly&utm_content=post) #Qwen (?q=%23Qwen) #AMD (?q=%23AMD) #local (?q=%23%D0%BB%D0%BE%D0%BA%D0%B0%D0%BB%D1%8C%D0%BD%D1%8B%D0%B5) #HALOGEN (?q=%23HALOGEN) #Fedora (?q=%23Fedora) #LLM (?q=%23LLM) #neuralnetworks (?q=%23%D0%BD%D0%B5%D0%B9%D1%80%D0%BE%D1%81%D0%B5%D1%82%D0%B8) #localAI (?q=%23%D0%BB%D0%BE%D0%BA%D0%B0%D0%BB%D1%8C%D0%BD%D1%8B%D0%B9%D0%98%D0%98)

🔗 Read on habr.com