Running Local LLMs on 285HX: CPU is Surprisingly the Fastest
nirurin · reddit · 2026-07-12
A user tested local LLM execution on an Intel 285HX mini PC, comparing different Llama.cpp backends and multiple models, focusing on speed differences among CPU, Vulkan, and SYCL+iGPU.
Test Environment
- Machine: MS-02, Intel Core Ultra 285HX, 64GB RAM
- Tool: Llama.cpp (tried Llama-Swap, but it wasn't very convenient under SYCL)
- Models: Various quantized versions of Qwen3, Qwen3.6, and Gemma4
Results
- Vulkan: Slowest overall, with some models hitting only 0.5–4 tk/s
- SYCL (iGPU+CPU): Significantly better than Vulkan, around 8–12 tk/s
- CPU-only: Surprisingly the fastest, reaching about 16.5 tk/s for Qwen3-30B and 14 tk/s for Qwen3.6-35B
Observations & Conclusions
- Past resources generally suggested Vulkan > SYCL, but on this generation of ARC-based iGPUs, the results appear reversed, or at least SYCL isn't as bad as the documentation implies.
- The author is unsure if the iGPU brings efficiency gains since no power consumption tests were done.
- Currently, pure CPU is the fastest solution.
- In these tests, Gemma4 isn't really worth it, while the CPU solutions for Qwen3 / Qwen3.6 are quite practical, with the author leaning towards recommending trying Qwen3.6 first.
Related event: Intel 285HX Tests Show CPU Beating iGPU for Local LLMs(2 posts)→
More from Infra
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27
- RTX Spark laptops impress analyst: 30B local model runs quiet with great battery life — Scobleizer · 2026-10-27
- Reflection AI's Beam-1.0 tops US open-weight models, hitting 88% of Qwen3.8-2.4T with 23B active params — ryanshrout · 2026-10-27
- Re-uploading a 239GB robot dataset moved only 2.6MB — inside a petabyte-scale data pipeline — mervenoyann · 2026-10-27
- Intel to acquire Classiq to build full-stack quantum computing platform — BenBajarin · 2026-10-27