XeStrata Runs Qwen3.8-Flash-Next on Intel, AMD, and NVIDIA GPUs
Developer MistVVK released XeStrata, free open-source software that runs the 125-billion-parameter Qwen3.8-Flash-Next model on ordinary PCs.
1 source 2h ago
Reporting from 1 source: GIGAZINE.
Why it matters
XeStrata moves local execution of a 125-billion-parameter model off CUDA and HIP and onto a single SYCL code path, which lowers the vendor lock-in that has shaped which GPUs can run large models at home.
Key Facts
- XeStrata is free open-source software from developer MistVVK that runs the 125-billion-parameter Qwen3.8-Flash-Next model on ordinary PCs.
- XeStrata is based on Strata version 0.1.24 and rewrites GPU compute using the Khronos Group's SYCL standard.
- The free build supports Intel and AMD GPUs using only open-source software; the contrib-llvm build adds NVIDIA GPUs and can use Intel's oneMKL and NVIDIA's cuBLAS.
- XeStrata uses XMX on Intel GPUs, WMMA on AMD GPUs, and Tensor Core on NVIDIA GPUs, and also supports NVIDIA's pre-Tensor-Core Pascal generation at lower speed.
- XeStrata is released under the LGPL-3.0 or later license, and code inherited from Strata keeps Strata's MIT license.
It is based on Strata version 0.1.24 but rewrites GPU compute with the Khronos Group's SYCL standard, so Intel, AMD, and NVIDIA GPUs share common code. A free build covers Intel and AMD GPUs using only open-source software; a contrib-llvm build adds NVIDIA support and manufacturer compute libraries.
XeStrata forks Strata 0.1.24, the software that splits a model's data across GPU VRAM and main memory so a large model fits on a desktop. The fork keeps that mechanism and replaces the compute layer with SYCL, the Khronos Group standard behind OpenGL and Vulkan. Each vendor's dedicated units are still used: XMX on Intel, WMMA on AMD, Tensor Core on NVIDIA. NVIDIA's pre-Tensor-Core Pascal generation also runs, at lower speed than GPUs with those units.
The free build supports Intel and AMD GPUs with open-source software alone. The contrib-llvm build adds NVIDIA and can use Intel's oneMKL and NVIDIA's cuBLAS. XeStrata is under LGPL-3.0 or later; code inherited from Strata keeps its MIT license.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google