Effective harnesses narrow bug-finding gap for open-weight models
The author compares several open-weight LLMs (DeepSeek V4, Qwen3.5, Kimi, GLM-5/5.1) against the closed-source Opus 4.7 on the crackaddr vulnerability suite. While open models lag on tougher artifacts, a well-designed harness and post-training tweaks close most of the gap, with GLM-5.1 matching Opus performance.