小语言模型(SLM, Small Language Model)是参数量在数十亿以下、可在消费级或边缘设备本地运行的紧凑语言模型,核心是以专注任务换取轻量高效,能力包括低显存推理、离线运行与快速响应。它解决的是大模型成本高、延迟大与数据出域的合规风险,并不解决开放域复杂推理的能力上限问题。
小语言模型的核心是够用且可控,它可以帮助企业实现以下目标:
让端侧场景落地。手机、POS 机与工业网关上直接运行,断网与弱网环境照常提供智能服务,业务连续性不再系于云端。蒸馏与量化技术让小模型在大模型教师指导下获得成比例的能力迁移,「小而专」的路线有了成熟的方法论。任务路由网关按复杂度自动分发请求,小模型承接大部分流量、大模型处理疑难问题,成为性价比最优的混合架构。
让敏感数据不出域。本地推理满足金融、医疗等行业的数据合规硬约束,隐私审批流程大幅简化。端侧推理框架对主流芯片的适配日趋完善,从手机 NPU 到车机平台都有可用的部署方案。
让成本结构改善。高频简单任务(分类、抽取、格式转换)由小模型承接,整体推理账单明显下降,算力预算更从容。
让响应更即时。本地毫秒级首字延迟,交互体验摆脱网络波动影响,实时场景的用户满意度提升。
总结来看,小语言模型与大模型不是替代关系而是分工关系,「合适的任务用合适大小的模型」是 AI 工程化的成熟思路。
翻译
A Small Language Model (SLM) is a compact model, typically below tens of billions of parameters, that runs locally on consumer or edge devices, trading breadth for efficiency with low-memory inference, offline availability and fast responses.
The core of SLMs is being sufficient and controllable, and they help enterprises achieve the following goals:
Edge scenarios land: phones, POS terminals and industrial gateways run models directly, keeping service alive without the cloud.
Sensitive data stays local: on-device inference satisfies hard compliance constraints in finance and healthcare.
Better cost structure: high-frequency simple tasks are offloaded to small models, noticeably lowering inference bills.
Instant responsiveness: local millisecond first-token latency frees user experience from network jitter.
In summary, SLMs complement rather than replace large models — using the right size for the right task is mature AI engineering.