XeMaya brings native DeepSeek V4.1 and GLM 5.3 inference to Intel Arc GPUs
Jumbomuffin1312 · reddit · 2026-11-11
Reddit user Jumbomuffin1312 released XeMaya, a fork of XeStrata that runs newer MoE models like DeepSeek V4.1 Flash and GLM 5.3 Flash natively on Intel Arc GPUs via SYCL and Level Zero.
Benchmarks (prefill/generation tok/s, author cautions these aren't strictly controlled):
- GLM 5.3 Flash: stock llama.cpp SYCL hits 68.3/13.6; XeMaya reaches 152–161/17.4 (20 generation with MTP)
- DeepSeek V4.1 Flash Coder: 48.5/16.3 initially, 224.8/21.7 after optimization
- DeepSeek V4.1 Flash full 384-expert model: patched llama.cpp managed 45.6/10.9; XeMaya only 11.4/14.56, and two GPUs beat three because the B60 sits on a slower chipset link
- Qwen3.8 Flash-Next: XeMaya 1,842–1,884/76.6, still behind Strata's 2,211/78.3
Tested on an Ubuntu server with B70 32GB + B65 32GB + B60 24GB, a Ryzen 9950X, and 128GB RAM. Early alpha, MIT-licensed on GitHub, crediting XeStrata, Strata, Project Maya, and the llama.cpp community.
Related event: Open-source XeMaya engine runs DeepSeek V4.1 on Intel Arc GPUs(2 posts)→
More from Infra
- Claude ported itself to AMD MI355X in ~3 days, work that took dozens of engineers 6-12 months — AnushElangovan · 2026-11-11
- Analyst: AI datacenter demand will eventually be built in space — TinfoilTricorn · 2026-11-11
- Local LLMs are bare engines, cloud APIs are finished cars: what newcomers miss — Oli_Luck · 2026-11-11
- Proposal: build data centers underground and use waste heat to melt road snow — MatthewChang · 2026-11-11
- SpaceX to build $24B AI data center and natural gas power plant in Starbase, Texas — Polymarket · 2026-11-11
- Microsoft's DLSS rival Auto SR launches on Xbox Ally X with handheld 1080p and 120fps support — tomwarren · 2026-11-11