01
Apple M4 Pro · 48GB7B–32B Chinese models benchmarked
Comparing time-to-first-token, generation speed, memory and long-context behavior.
Beyond leaderboards — reproducible
Open models, quantizations, engines, devices, parameters, scripts and raw data. Every conclusion states its limits.
02
RTX 4090 · 24GBLocal coding models and agent task evaluation
From code completion to tool calling, multi-file edits and task recovery.
03
CPU ONLYNo GPU — what can local AI do?
Covering 8 common desktop CPUs with model limits and acceptable experience ranges.
04
ANDROIDSpeed, power and thermals of small on-phone models
Comparing on-device engines and quantization schemes under unified tasks and conditions.