作者 | 平章
出品 | 《易现场》
2026年,大模型竞赛进入“下半场”。当互联网巨头仍在比拼云端模型的参数规模与推理速度时,手机厂商却正在悄然启动一场“端侧革命”。
9月16日,2026 vivo开发者大会现场,vivo副总裁、OS领域副总裁、vivo AI全球研究院院长周围向《易现场》等阐述了端侧大模型、平台底座与个人化智能的难题。
相较往年“不盲目追参数量”的保守口径,这次周围明确表态vivo已开始预研蓝心30B MOE端侧模型,并把“专家稀疏加载”作为手机跑大模型的核心解法。在他看来,端侧大模型不是“能不能跑”的难题,而是“为什么必须跑”的战略必选项——隐私控制权、个人化智能、平台级执行,以及芯片硬化带来的功耗优化,共同构成了手机厂商必须走端侧的核心动因。
这不仅是一次工艺释疑,背后更是一家头部手机厂商对“为什么必须走端侧”的深层思考。
为什么从“全量加载”转向30B MOE端侧模型
周围一开场就坦言,去年vivo对端侧大模型的判断偏谨慎:7B、3B全量载入已对内存形成压力。
周围复盘了这一转变:“好家伙,在今年年初的时候发现,大模型实际上不是要一次就加载完的,它是可以MOE,每次只要加载几个专家模型就行了。”在他看来,手机本身具备512GB甚至1TB的存储空间,旗舰机内存也普遍来到12GB,拿出2-3G内存运载MOE模型,让端侧跑大模型变成了现实。
按vivo当前预研口径,30B并非“稠密模型全跑”,而是总参数切分为数十到上百个专家,单次推理动态激活1-3B,内存占用可压到2-4GB。
“30B给人们的错觉是很大,实际上只是把它切成了30块或者100块。”周围解释道。作为参照,Qwen3-30B-A3B总参约30.5B、激活约3.3B,在桌面全量量化仍常需20GB级内存;vivo的2-4GB属于高度量化、按需分页、只常驻部分专家的端侧工程口径。
周围坦言,30B MOE当下仍是工程验证,不会立刻上量产机型。原因有两层:模型本身需调校,硬件需等芯片规划,“今年提液态玻璃是因为24个月前就规划了芯片,今年提MOE,同样需要24个月后对应的芯片规划落地。”
他预测,到2028年底,MOE达成大模型能力在端侧一定会变成现实。
功耗与常驻:端侧该如何“跑”?
端侧30B最容易被质疑的是电池续航。对此,周围向《易现场》做出阐述:专家常驻RAM本身耗电极低,真正耗电在于“搬运”--从存储或闲置专家区调参进计算单元。
据周围透露,3B级初期思路工作电流曾约800mA,迭代后可降到500-600mA;30B MOE若只切专家、由算力芯片完成搬运与推理,理想工作电流约500mA,而2028年其工程目标为600mA上下。
这要求平台层做三件事:其一是统一渲染与计算资源调度。OriginOS 7的通透玻璃、动态光影并非纯软达成,2024年vivo已把高斯模糊、图层合并硬化进芯片,避免走CPU软绘带来的功耗崩塌;其二是模型权重分层,高频专家常驻、低频专家存储待命,跨域任务只换专家不换全模型;其三是端侧推理与OS Harness融合,蓝心小V调用相册、日历、设置等本机能力时不必每次传云。
周围以相册为例,“端侧模型后台默默感知并记忆了特征,使用者问‘国庆带小孩在海边’的照片,瞬间就能完成。相册100GB照片如果在服务器跑要半天,但端侧模型在Main-Loop里早就建好了个人化存储空间。”这类感知记忆,正是30B MOE后续替代小模型、做个人化编排的基础。
手机厂商为什么一定要走端侧?
把30B MOE放在vivo总体AI战略里,把Agent从“语音触发—云上推理—返回结果”改成了“本地理解—本地预判—人工确认”。
为什么手机厂商必须走端侧?以下是vivo高管们强调的核心逻辑。
第1, 隐私与数据控制权。周围多次强调,手机是贴身设备,相册、聊天记录、家庭关系等敏感数据若全部上云,隐私风险不可控。而鉴于vivo秉承着“使用者是机场塔台,飞机起不起飞由使用者说了算”的安全原则,可在本地完成感知、记忆和推理的端侧模型,数据不出端,使用者可一键删除所有AI记忆,这就是vivo坚持自研端侧模型、而非完全依赖云的关键因素。
第2, 个人化智能需要本机全量上下文。vivo AI机型总经理关岩冰指出,人们用各种Agent时总要反复输入背景信息,而手机上存着使用者最全的个人数据。
周围举了一个生动的例子:做旅游攻略,第三方APP也能做,但如果你老婆怀孕了,你可能忘了跟第三方APP说;而蓝心小V会很贴心地说“能不能带个折叠小凳子?”还有,你看了《飞驰人生》聊了巴音布鲁克,小V去新疆就会建议你去巴音布鲁克。“只有懂你家庭关系、社会关系和行为偏好的Harness,才能做出让你尖叫的机型。”这种个人化,必须依赖端侧长期记忆和本地模型。
第3, 平台级执行与原子能力调用。vivo当前已把本机6000个原子能力接入小V,长期目标按PC Harness标准补足到约3万个接口。30B MOE可作为本机编排大脑,直接调用设置、相册、日历、通信等,不必所有步骤走GUI自动化。
周围直言:“即使在PC上,私有应用对GUI的兼容也格外弱,手机不需要那么执着。”端侧模型可以直接理解意图、调用底层能力,这是第三方云助手无法做到的,它们没有平台权限。
第四,芯片硬化与功耗优化的终端厂商长处。纯软算法无法达成液态玻璃成效,vivo在2024年就把高斯模糊硬化进芯片;同样,端侧大模型的低功耗运行必须依赖NPU、内存带宽和算力的协同外观。周围表示,只有终端厂商去和芯片厂商合作,才能把这个端侧做得比较好。这也是为什么vivo坚持自研模型并走端侧路线。
第五,无网络情形与无障碍刚需。周围特意表示,vivo的无障碍能力(如vivo听见)在飞机上无网络时,仍旧可以全程端侧语音识别,服务听障人群。这证明端侧模型不是噱头,而是真实情形的必选项。
基于以上原因,vivo判断:将来Agent手机的核心竞争力不在于云端基模有多大,而在于端侧模型能否与平台深度融合,形成个人化智能。
采访最后,周围再次强调了vivo的机型哲学:“机型的竞争力不是和友商比,而是是否真正让使用者满意。就像无障碍能力,笔者做出了让使用者偏爱的、好很多的机型。将来,蓝心小V会成为你专属的个人助理,这是笔者的愿景。”
本文来源:https://www.163.com/tech/article/L70UUPDH00097U7T.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
0
Author | Ping Zhang
Publisher | 《Yi Live》
In 2026, the race for large models enters its “second half.” While internet giants are still competing over cloud model parameter scale and inference speed, smartphone manufacturers are quietly launching an “edge computing revolution.”
On September 16, at the 2026 vivo Developer Conference, Zhou Wei, Vice President of vivo, Vice President of OS, and Director of vivo AI Global Institute, explained to 《Yi Live》 and others the challenges of edge-side large models, platform foundations, and personalized intelligence.
Compared to last year’s conservative stance of “not blindly chasing parameter counts,” Zhou Wei clearly stated this time that vivo has begun pre-research on the Blue Heart 30B MOE edge-side model, treating “expert sparse loading” as the core solution for running large models on smartphones. In his view, the issue with edge-side large models is not “whether they can run” but “why they must run” — privacy control, personalized intelligence, platform-level execution, and power optimization from hardware hardening collectively make the edge side a strategic necessity for smartphone manufacturers.
This is not just a technical clarification; it represents a leading smartphone manufacturer’s deep reflection on why edge computing is indispensable.
Why Shift from “Full Loading” to 30B MOE Edge Model
Zhou Wei began by admitting that vivo was initially cautious about edge-side large models last year: fully loading 7B or 3B models already strained memory.
He recounted the turning point: “In early this year, we discovered that large models don’t need to be loaded all at once — they can be partitioned into MOE layers, where only a few expert modules are loaded at any given time.” With smartphones now having 512GB or even 1TB storage and flagship models boasting 12GB RAM, allocating 2–3 GB for MOE loading makes running large models on edge a reality.
According to vivo’s current pre-research, the 30B model isn’t fully loaded as a dense net. Instead, it’s split into dozens to hundreds of experts, activating only 1–3B parameters during inference — keeping memory use under 2–4 GB.
“30B suggests bigness, but it’s actually sliced into 30 or 100 pieces,” Zhou explained. For context, Qwen3-30B-A3B (with ~30.5B total, ~3.3B active) still needs ~20 GB RAM in desktop setups. Vivo’s 2–4 GB budget assumes heavy quantization, lazy loading, and sparse activation — hallmarks of edge engineering.
Zhou confirmed that 30B MOE is still in experimental phase and won’t ship on production models yet. The reasons are twofold: the model needs fine-tuning, and hardware must catch up — “Announcing liquid glass last year followed 24 months of chip planning. Mentioning MOE now means we’re preparing for the post-2027 hardware wave.”
He predicts that by late 2028, MOE-based edge large models will be commonplace.
Power Consumption and Residency: How to “Run” on Edge
A common critique of edge-side 30B is battery drain. Zhou clarified: while resident experts consume little energy, the bulk comes from “transferring” them — loading parameters from storage into compute units.
Early implementations for 3B models drew ~800 mA, but optimization dropped it to 500–600 mA. For 30B MOE, where only active experts are streamed and computed by an NPU, the target current is ~500 mA — achievable by 2028 with better hardware.
This requires three platform innovations: first, unified rendering and computation scheduling — as seen in OriginOS 7’s liquid glass effects, which rely on GPU-hardened blur and layering rather than CPU emulation; second, hierarchical weight management — keeping frequently used experts hot in RAM and lazy-loading rare ones; third, deep integration of inference with the OS harness, enabling local execution of tasks like photo tagging or calendar queries without sending data to the cloud.
Zhou illustrated this with a photo example: “An edge model silently learns features during idle time, so when you ask for ‘Labor Day beach shots with kids,’ it retrieves them instantly. If this ran on a server with 100 GB of photos, it’d take ages. But on device, we can maintain a personalized index continuously.”
Why Must Smartphone Vendors Go Edge?
The shift from “voice trigger → cloud inference → result return” to “local understanding → local prediction → human confirmation” hinges on the edge model’s presence.
Here are the core reasons why smartphone makers must embrace the edge:
1. **Privacy & Data Control**. As Zhou emphasized, phones are intimate devices storing sensitive data — photos, chats, family records. Sending them to the cloud risks privacy breaches. Vivo’s approach — “Users are the control tower; they decide when the plane flies” — ensures local execution keeps data secure. Users can erase all AI memory with a tap, which isn’t possible solely relying on cloud services.
2. **Personalized Intelligence via Full Context**. Guan Yanbing, General Manager of AI Devices at vivo, noted that repeated context insertion frustrates users interacting with agents. Smartphones contain the richest personal datasets — but only edge models can access and exploit them fully.
Zhou gave a vivid example: during travel planning, a cloud service might miss that your wife is pregnant, but Blue Heart will suggest packing a folding chair. Or if you discussed “Flying in the Wind” and Bayinbuluk, it’ll recommend that stop on your Xinjiang trip. This level of personalization demands long-term local memory and context-sensitive processing.
3. **Platform-Level Execution & Atomic API Access**. Vivo currently exposes ~6,000 atomic APIs to its assistant but aims for ~30,000 by PC-level standards. A 30B MOE on edge could orchestrate these directly — adjusting settings, querying calendars, or managing communications — without relying on fragile GUI automation or third-party cloud gateways.
“Even on PCs, proprietary apps struggle with GUI compatibility,” Zhou said. “Mobile doesn’t need that overhead. We can map intent to system APIs directly.”
4. **Hardware Hardening & Efficiency**. Only terminal vendors have the leverage to harden algorithms into silicon. Vivo’s success with liquid glass relied on offloading Gaussian blur to the GPU; similarly, running large models efficiently requires tight collaboration between software and chip designers to minimize latency and power.
5. **Offline Reliability & Accessibility**. Zhou highlighted that vivo’s accessibility features — like real-time speech recognition for the hearing impaired — must work offline, e.g., during flights. This proves edge models aren’t optional — they’re essential for inclusivity and reliability.
In short, future smartphone competitiveness won’t lie in cloud model size, but in how well edge-side models integrate with the platform to deliver personalized, private, and robust intelligence.
“Our goal is simple,” Zhou concluded. “Blue Heart should be your personal assistant — remembering your likes, protecting your data, and making life easier. That’s why we’re betting on the edge.”
Source: NetEase Intelligence; Editor: Wang Fengzhi_NT2541


