Mini server test shows slow local AI without a dedicated GPU

A firsthand MS-02 mini PC homelab test tried running without a dedicated graphics card. The machine used an Intel Core Ultra 285HX processor and 64 GB of memory, mostly through the default Docker releases of Llama.cpp. Llama-Swap made quick model switching useful, but it did not work well with SYCL, which made testing harder.

Vulkan was tested with iGPU access enabled through /dev/dri, but the system still seemed to use a lot of CPU at the same time instead of moving the work only to the iGPU. Qwen3-30B-A3B-Instruct-2507-IQ4_NL ran, but only at about 2 . Two versions, Q4_K_S at 4.22bpw and IQ4_XS at 3.93bpw, each reached about 0.5 .

Gemma 4 26B A4B MXFP4 MOE was also part of the test, but the available details do not include a complete speed result.

Key points

  • The test machine was an MS-02 mini PC with an Intel Core Ultra 285HX and 64 GB of memory.
  • Most testing used Llama.cpp Docker builds with default settings.
  • Llama-Swap helped with fast model switching, but SYCL problems made it less useful here.
  • Vulkan with iGPU access did not appear to remove heavy CPU use.
  • Qwen3 30B reached about 2 , while tested Qwen3.6 35B versions were around 0.5 .

Sources covering this story (2)

Read original