The author tested 14 local LLMs on a GMKtec EVO-X2 mini-PC equipped with 96 GB of RAM. The system runs on Ubuntu 26.04, utilizing LM Studio and Q4 quantization to ensure models fit into memory. The testing process spanned 86 hours of runtime to evaluate performance across various tasks.
The evaluation methodology consisted of two main stages: a 'ladder' test involving 39 discipline-based tasks, and a 'practice' test consisting of six applied problems. The author measured performance based on instruction following, tool calling capabilities, and reasoning speed, noting that context window size did not significantly impact speed or accuracy.
Performance results varied by model architecture. Dense models like Qwen3.8 27B achieved approximately 12.8 tokens per second, while MoE models like Gemma 4 26B reached up to 63 tokens per second. The author observed that long context processing times ranged from 8 to 23 minutes for 200K tokens.
Following the tests, the author selected two models for permanent use. Gemma 4 26B Uncensored is utilized for conversational tasks, while Qwen3.8 27B handles complex assignments and background processing. The author plans to continue refining the testing methodology as new models are released.
