四款模型，一个架构，端侧AI的分层计算逻辑终于被手机厂商想明白了。
vivo在2026开发者大会上发布了四款核心蓝心大模型：BlueLM-RealTime、BlueLM-Nano、BlueLM-Flash和BlueLM-Pro。这不是简单的"大中小"分级——四款模型各有明确的能力定位和部署场景，共同构成了一套从端侧到云端的分层AI架构。
BlueLM-Nano：3B参数端侧落地
四款模型中最具突破性的是BlueLM-Nano。这款模型仅有30亿参数，但首次实现了在手机端侧的完整部署——包括自然语言理解、多轮对话、本地知识库检索等核心能力。3B参数是端侧大模型的一个关键门槛：再小则能力不足，再大则功耗和存储压力过大。vivo选择在3B参数上实现端侧落地，说明他们在模型压缩和推理优化上已经找到了实用化的平衡点。
更具前瞻性的是，vivo同时披露了正在预研的30B MoE（混合专家）端侧大模型。MoE架构的核心优势是：模型总参数量很大，但每次推理只激活部分"专家"子网络，实际计算量远小于同等参数的稠密模型。如果30B MoE能在端侧跑通，将意味着手机本地就能处理相当复杂的AI任务，而无需依赖云端。
原系统7 + 蓝心Harness：AI作为系统能力
vivo同步发布了原系统7，将蓝心大模型深度集成到操作系统层面。蓝心Harness作为AI能力的调度中枢，负责在不同场景下自动选择最合适的模型——实时响应用RealTime，轻量任务用Nano，复杂推理用Pro。用户感知不到模型的切换，只感受到"手机变聪明了"。
BlueCode：编程智能体入场
vivo还发布了BlueCode编程智能体，面向开发者的AI辅助编程工具。这意味着vivo的AI战略已经从"提升终端用户体验"扩展到了"赋能开发者生态"。当手机厂商开始做编程智能体，说明AI竞争的边界正在从消费者端向开发者端延伸。
四款模型、一个操作系统、一个编程智能体——vivo正在构建的不是单一的AI产品，而是一套从底层到应用的全栈AI基础设施。端侧AI的竞赛，才刚刚开始。
明天见。
Four models, one architecture. On-device AI's layered compute logic has finally been figured out by a phone maker.
At its 2026 Developer Conference, vivo unveiled four core BlueLM models: BlueLM-RealTime, BlueLM-Nano, BlueLM-Flash, and BlueLM-Pro. This isn't simple "small-medium-large" tiering — each model has a clear capability profile and deployment scenario, together forming a layered AI architecture from on-device to cloud.
BlueLM-Nano: 3B Parameters On-Device
The most breakthrough of the four is BlueLM-Nano. At just 3 billion parameters, it achieves full on-device deployment for the first time — including natural language understanding, multi-turn dialogue, and local knowledge base retrieval. 3B parameters is a critical threshold for on-device LLMs: smaller and capability is insufficient; larger and power/storage pressure becomes unmanageable. vivo's choice to land at 3B signals they've found a practical balance between model compression and inference optimization.
More forward-looking: vivo simultaneously disclosed a 30B MoE (Mixture of Experts) on-device LLM in pre-research. MoE architecture's core advantage: total parameters are massive, but each inference only activates a subset of "expert" sub-networks, making actual compute far smaller than an equivalent dense model. If 30B MoE runs on-device, it means phones handle significantly complex AI tasks locally — no cloud required.
OriginOS 7 + BlueLM Harness: AI as a System Capability
vivo also launched OriginOS 7, deeply integrating BlueLM into the OS layer. BlueLM Harness serves as the AI capability dispatcher, automatically selecting the optimal model for each scenario — RealTime for instant responses, Nano for lightweight tasks, Pro for complex reasoning. Users don't perceive model switching — they just feel "the phone got smarter."
BlueCode: The Coding Agent Enters
vivo also released BlueCode, an AI-assisted coding tool for developers. This means vivo's AI strategy has expanded from "improving user experience" to "empowering the developer ecosystem." When a phone maker starts building coding agents, AI competition's boundary is extending from consumer-facing to developer-facing.
Four models, one OS, one coding agent — vivo is building not a single AI product, but a full-stack AI infrastructure from bottom to top. The on-device AI race has only just begun.
See you tomorrow.
端侧AI的竞赛不是比谁的模型更大，而是比谁先把大模型变成用户感知不到的系统能力。
—— Dawn Vision编辑部
The on-device AI race isn't about who has the biggest model — it's about who first makes LLMs invisible as a system capability.
— The Dawn Vision Editorial Desk
Dawn Vision, vivo, BlueLM, on-device LLM, 3B parameters, MoE, OriginOS 7, BlueCode, developer conference, AI commercialization
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的 10 个源信号生成，经编辑部人工审核。素材来源：IT之家、36氪、太平洋科技。
This article was generated by the Dawn Vision cognitive engine processing 10 source signals, with human editorial review. Sources: IT之家, 36Kr, PConline.
