A Local 27B Model Tied Sonnet on Coding. 300 Benchmarks Show Where It Actually Wins.
Everyone has an opinion about whether local models can replace cloud models for coding. I ran 300 benchmarks on Qwen3.8-27B against a corpus of 212,000 Claude runs. It tied Sonnet. Only two of eight comparisons survived the confidence intervals, and one of my own findings turned out to be a bug in my harness.
Read more →