XeMaya brings native Intel Arc inference to recent MoE models
A developer built XeMaya, a fork of XeStrata that runs models natively on Intel Arc GPUs through SYCL and Level Zero. The early-alpha project reportedly supports DeepSeek V4.1 Flash and Coder, GLM 5.3 Flash, and Qwen3.8 Flash-Next, with preliminary prefill and generation results reported on an Ubuntu server. The developer cautioned that different model files, prompts, and GPU configurations made the comparisons uncontrolled, leaving further fixes and optimization needed.
XeMaya extends native Intel Arc inference options to several recent MoE models, although its reported performance still requires controlled, independent validation.
I've been trying to get some of the newer MoE models running on my Intel Arc GPUs, mainly DeepSeek V4.1 Flash and GLM 5.3 Flash. Intel support for these models is still pretty limited, and while I could get some things working through patched llama.cpp builds, performance wasn't exactly great.
fable 5.2 ended up making XeMaya, a fork of XeStrata focused on getting these models running natively on Intel Arc through SYCL and Level Zero. The idea is basically to bring the kind of model support Project Maya has to Intel GPUs.
I've been testing it on my Ubuntu server with a B70 32GB, B65 32GB, B60 24GB, Ryzen 9950X, and 128GB RAM.
Here are some of the results so far. All speeds are prefill / generation in tok/s.
GLM 5.3 Flash
- Stock llama.cpp SYCL: 68.3 / 13.6
- Experimental grouped-matmul SYCL: 108.9 / 14.0
- XeMaya: 152–161 / 17.4 (around 20 generation with MTP)
#XeMaya#XeStrata#Strata#Project Maya#Intel Arc#DeepSeek V4.1 Flash#DeepSeek V4.1 Flash Coder#GLM 5.3 Flash#Qwen3.8 Flash-Next#llama.cpp#fable 5.2