Qwen3.5-9B on MacBook M5 Pro Scores 93.8% on HomeSec-Bench — 4.1pts Behind GPT-5.4

SharpAI·medium signal

SharpAI's HomeSec-Bench (96 tests across 15 security assistant workflow suites) shows Qwen3.5-9B running locally on a MacBook Pro M5 achieves 93.8% pass rate at 25 tokens/second, using only 13.8GB unified memory and zero API cost — versus GPT-5.4's 97.9%. The local model outperforms GPT-5.4-nano (92.7%) and nearly matches GPT-5.4-mini (95.8%). The larger Qwen3.5-35B-MoE delivers faster first-token response (435ms) than all OpenAI cloud models tested. The story got 160 HN points and 147 comments, confirming sustained practitioner interest in local AI matching cloud capability on specialized tasks.