SharpAI Aegis benchmark pits Qwen3.5, GPT-5.4

This is an eye-catching local-AI result, but it’s also a very opinionated benchmark built for one product’s security workflow. The real takeaway is less “open models beat frontier APIs” and more “a tuned 9B model is now plausible for serious on-device security automation.”

• –The strongest number is practical, not theoretical: 13.8 GB unified memory and 25 tok/s make the setup believable on a high-end laptop.
• –HomeSec-Bench tests workflow skills that matter for a camera/security agent, like event deduplication, prompt-injection resistance, alert routing, and JSON compliance.
• –The comparison is useful, but it’s not neutral: the benchmark is created around SharpAI Aegis, so treat the GPT-5.4 gap as directional rather than universal.
• –The 35B MoE result is also interesting: lower TTFT than the cloud models hints that local inference can be competitive on responsiveness, not just cost.
• –For buyers, the real appeal is privacy plus zero API spend, which is a compelling combo for always-on home monitoring.