Qwen3.5-9B has achieved impressive results on the new HomeSec-Benchmark, scoring 93.8%—within 4 points of GPT-5.4—while running entirely on a MacBook Pro M5. The model processes 25 tokens per second with 765ms TTFT, using just 13.8GB of memory. This demonstrates local AI's potential to match cloud performance without API costs or privacy concerns.

HomeSec-Bench evaluates LLMs on actual home security assistant workflows, not generic chat. The benchmark includes 96 tests across 15 suites covering security classification and tool use. Qwen3.5-35B-MoE even outperforms cloud models with 435ms TTFT versus 508ms for GPT-5.4-nano. All 35 fixture images are AI-generated, allowing testing against any OpenAI-compatible endpoint.

Called Aegis-AI, this local-first system runs on consumer hardware, offering complete privacy. The benchmark executes live on Apple Silicon with all tests visible in real time. Developers can download HomeSec-Bench from GitHub, which includes 96 LLM and 35 VLM tests across 16 suites. Consumer devices now handle sophisticated AI tasks traditionally requiring cloud infrastructure.