CPU-First Local Agent on 16GB RAM (DeepSeek Distill): Verifier-First, Audit-First Design

Abstract (EN) We present a CPU-first design blueprint for a local, tool-using conversational agent under a common consumer constraint: a modern CPU and 16GB RAM, with no always-on cloud inference. The central claim is that usefulness under tight compute is achieved less by maximizing raw model capability and more by constraining system degrees of freedom: structured IO, retrieval-as-evidence, verifier-first execution, and action-level gating for irreversible operations. We specify system modules, operational metrics, a minimal benchmark protocol, and a reference demo implementation. For the base model family, we recommend DeepSeek distillation-based instruct models (CN/EN capable) and a narrow adaptation surface (LoRA/DPO) focused on routing, schema adherence, and safety-critical decisions rather than general knowledge acquisition. 摘要(中文/Chinese) 本文提出一个在"CPU + 16GB 内存 + 本地离线"为约束条件下,构建可用思维对话/工具代理系统的工程化蓝图。核心观点不是追求"更聪明的模型",而是通过结构化输入输出、证据检索(retrieval-as-evidence)、可验证执行(verifier-first)以及不可逆动作门控(action gate)来降低系统自由度,从而在小模型上获得稳定、可审计的可用性。文中给出模块化架构、可操作指标与最小评估协议,并提供可复现的参考 demo。模型选型上建议以 DeepSeek 蒸馏系指令模型为底座(中英双语),把微调面收敛到路由、schema 遵循与安全关键决策上,而非在本地重新学习通用知识。