手机里的 AI 已经能写总结、修照片、陪你聊天,可当你想真正把一件事交给它——安排一次出差、跟进一个项目、订下一场聚会——它常常又把难题原样还给你:我好像不明白。
2026 年,AI 正在从「回答难题」走向「执行任务」,而真正的智能体,不会只活在某一台设备里:
电脑上开好的头,手机应该接得上;耳机听到的吩咐,手机应该办得成。
设备各有各的聪明,人却还得负责穿针引线——这正是高通想应对的难题。
今年骁龙峰会,高通更新了手机、PC 和音频平台,微软、谷歌等伙伴也带来了新机型。

看似分散的发布,指向同一个目标:AI 想包办你的一切,高通想包办你的 AI。
至尊之上,还有超级至尊
今年,骁龙手机旗舰的名字又长了一点。
第六代骁龙 8 至尊版与第六代骁龙 8 超级至尊版同时登场。
前者面向更广泛的旗舰终端,后者代表高通在表现、AI、游戏和影像上的最高规格。两者使用不同的封装工艺,超级至尊版尤其强调高负载下的持续表现。

制程上,两款芯片都基于台积电 2nm(N2P)工艺,是高通首款 2nm 手机芯片,有着迄今为止高通骁龙手机芯片最强的表现。
最显眼的数字,是高达 5GHz 的主频。
超级至尊版的 Oryon CPU 超级内核最高达到 5GHz,表现内核最高达到 4GHz。高通称,这是首款最高主频达到 5GHz 的移动 CPU。

不过,这次与主频一起受到重视的,还有 FlexCache 缓存架构。
核心跑得再快,数据没到,也只能原地干等。
FlexCache 允许不同核心共享一个动态分配的缓存池。超级内核遇到重任务,可以调用几乎整个缓存池;任务转交给其他核心,相关数据也能留在缓存中,减少重新访问平台内存的次数。
智能体恰好需要频繁开展这样的交接。一句出行安排,背后可能涉及读取日程、查询路线、比较时间和调用服务。不同步骤由不同核心与模型接力,FlexCache 减少的,就是接力过程中的等待。

GPU 的变化同样围绕 AI 展开。
新增的 Adreno Matrix Cores 将专用 AI 计算核心引入图形管线,配合 Adreno Neural Fusion 完成超分和帧生成。18MB HPM 片上显存则保存中间数据,减少往返平台内存的开销。这套工艺已兼容 Unity 和虚幻引擎,高通还计划与 18 款手游合作,基本上你喜欢玩的都涵盖在内了。影像维度,视频规格提高至 8K 60fps 和 4K 240fps超高清慢动作,ISP 与 NPU 的协作进一步深入。
Hexagon NPU 也在缩短数据与计算之间的距离:共享内存扩大 50%,新增面向 Transformer 的 Element Accelerator,并兼容 300 亿以上参数的 MoE 模型。
芯片表现在往上走,散热也是个要应对的大难题。
Offset PoP 将封装内存移到一侧,为处理器区域留出专用散热路径。

从缓存到封装,这些变化都在回应同一件事:手机既要继续拍照、打游戏,也要承担智能体带来的新任务。
高达 5GHz 的主频,提高了算力上限,数据访问和散热效率,也决定了这份算力能发挥多少、维持多久。
毕竟,旗舰表现不能只有「三分钟热度」。
电脑这边,微软和谷歌都来了
个人 AI 要真正「上班」,得先从电脑开始。
办公是智能体最先落地的情形:读资料、拆目标、开软件、查结果,任务足够复杂,又足够通用——旅行规划、家庭事务需要的,实际上是同一套能力。
所以今年的骁龙峰会,电脑和手机一样站上了主角位。
微软带来的,是熟悉的 Surface:12 英寸 Surface Pro 和 13 英寸 Surface Laptop 搭载基于 3nm 工艺的骁龙 X2 Plus,给出亮铂金、罗兰紫和典雅黑配色。两款机型均提高了显示屏亮度,部分 Surface Pro 商用版本可选配 5G。
Surface Pro 还加入了 Ink Canvas,让 AI 参与触控笔创作。同期亮相的 Surface 鼠标兼容触觉反馈和自定义操作按钮。这批新品计划于 10 月 13 日起上市。

谷歌带来的,则是一种新的电脑:Googlebook。
它以 4nm 工艺的骁龙 X Elite 为核心,将 Android 生态带入笔记本,并针对 Gemini Intelligence 优化。高通披露了 19 小时以上的电池续航,原生应用和与 Android 手机之间的跨设备工作流,则是这条机型线的重点。
一边是 Windows 生态的延续,一边是 Android 生态的落地。两条路径殊途同归:都在骁龙平台上。

有意思的是,高通还官宣对 Linux 开源平台生态的兼容。在 AI PC 概念火热的当下,挑选 Linux 的使用者不少,他们也是高通的潜在客群。

高通的目标,是让基于 3nm 工艺的骁龙 X2 Elite 成为定义智能体 AI PC 的关键平台——骁龙 X2 系列给出最高 85 TOPS 的 NPU 算力,端侧 Token 生成速度最高达到前代的两倍,上下文兼容扩展至 20 万 Token,并计划展示超过 300 亿参数模型在本地运行。
这些指标对应着工作的不同需要。长上下文容纳资料与任务记录,本地推理减少对网络和云端服务的依赖。对于需要反复读取文件、连续修改内容的任务,模型既要算得动,也要记得住。
办公 Agent 之所以成为竞争焦点,也在于办公任务足够复杂,同时具有通用性。
读资料、拆解目标、操作软件、检查结果,这些能力同样适用于旅行规划和家庭事务。
办公 Agent 竞争的尽头,触及的正是通用 Agent 的能力。
电脑因而成为高通个人 AI 布局中绕不开的一站。

当下,骁龙 PC 兼容的模型已超过 450 个,Claude、Codex、OpenClaw、Perplexity 等应用已能在平台上运行,更多端侧 AI 开发商也加入了合作。
Surface 和 Googlebook 把芯片、软件与整机放到了一起。对高通而言,模型在本地跑起来只是开始;让这些能力进入每天打开的电脑、接下每天要做的工作,骁龙 PC 才有机会成为更多人的主力机。
AI 可以有想象力,办公电脑得有交付能力。
耳机,也想多管一点事
电脑合上之后,人们还有生活。
通勤时确认日程,走路时记下想法,购物时询问建议,这些需要零碎、即时,往往也说不完整。个人 AI 想参与其中,就需要知道此时此地发生了什么。
耳机、眼镜和手表的价值,由此显现。声音、第一视角、身体状态,都能补充显示屏之外的信息。设备贴近人,也就有机会理解那些没有被主动说出来的背景。

但贴身设备有个共同的难处:任务可以重,耳机眼镜可不能太重。
第二代骁龙音频平台至尊版接替此前的 S7/S7 Pro。相比前代,AI 表现最高提高至两倍,功耗最高降低 40%,占板面积减少 30%。对于耳机,后两项的关键性丝毫不亚于算力。

音乐、降噪和通话已经在消耗电量,AI 还要持续待命。新平台借助多核架构兼容这些任务并行运行,同时集成超低功耗 Wi-Fi,结合 XPAN 和蓝牙扩展连接范围,兼容访问云端智能与服务。
耳机无需独自完成所有计算,但要听得清、连得上。
高通给出了一个很具体的情形:戴着耳机在街上听歌,降噪原本应该隔绝环境;识别到逐步接近的警笛后,平台又会调整通透成效,让关键声音传进来。新的自适应 ANC 还会依据佩戴贴合度、运动和噪声变化调整。
降噪不能一味埋头苦干,耳机也得学会「听声辨事」。
再往前一步,它还可以接下具体事务。
高通推出的「骁龙畅听应用与智能体」生态,希望将应用和服务直接引入耳戴设备。MicroEJ 给出应用框架,使机型上市后仍能增加能力;Cinemo 等伙伴参与内容接入,Roon、Qobuz 也在其中。
比价、调整预约和安排出行,都是高通希望借助语音达成的感受。原本需要解锁显示屏、寻找应用的操作,有机会被一句话接替。

音质和降噪决定听觉感受,感知、连接与应用,则决定它在听歌之外还能做什么。耳机可以多管一点事,却不能多添一份麻烦:该提醒时开口,没事时安静,电量还得撑到回家。
从耳机到眼镜、手表,个人 AI 想进入通勤、运动和行走的间隙,都要应对相似的难题。算力可以藏在芯片里,存在感最好别压在脸上、挂在耳朵上。
过去,高通对手机的影响,最终体现为一张照片、一段视频,或一局流畅运行的游戏。消费者未必记得 ISP 和 GPU 的名字,却熟悉它们带来的感受。
个人 AI 也会走同样的路。耳机负责聆听,眼镜负责感知,手机和电脑负责思考与执行——设备各司其职,又彼此协同,AI 不再属于某一台设备,而是跟着人走。
到那时,使用者买的表面上是一台台硬件,实际挑选的,是一套关于自己的智能。

所以,手机、电脑和耳机的发布,讲的是同一门生意:模型可以来自不同厂商,平台可以各有性格,终端可以挂不同品牌,高通希望它们背后都有骁龙。
从手机到 PC,从耳机到眼镜、手表,每一条赛道都有高通芯片。个人 AI 的入口争夺战还在继续,谁也说不好下一个赢家会是手机、PC 还是耳机、眼镜。
而高通想做的,是让这些入口背后都是同一个骁龙。
本文来源:https://www.ifanr.com/1681946?utm_source=rss&utm_medium=rss&utm_campaign=
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
AI on smartphones can already write summaries, retouch photos, and keep you company in chats. But when you try to truly delegate a task—planning a business trip, following up on a project, or booking a party—it often just hands the difficulty right back to you: "I seem to not understand."
In 2026, AI is moving from "ansolving problems" to "executing tasks." And true agents won't just live on a single device:
What you start on a PC, your phone should be able to pick up; what you dictate to your earbuds, your phone should be able to execute.
Each device has its own intelligence, but people are still left to do the bridging—this is precisely the challenge Qualcomm wants to address.
At this year's Snapdragon Summit, Qualcomm updated its smartphone, PC, and audio platforms, while partners like Microsoft and Google unveiled new models.
These seemingly separate announcements point to the same goal: AI wants to handle everything for you, while Qualcomm wants to handle your AI.
Above the Extreme Edition, there’s a Super Extreme Edition
This year, the name of the flagship Snapdragon smartphone chip got a bit longer.
The Snapdragon 8 Extreme Edition Gen 6 and the Snapdragon 8 Super Extreme Edition Gen 6 debuted simultaneously.
The former targets a wider range of flagship devices, while the latter represents Qualcomm’s highest specs in performance, AI, gaming, and imaging. The two use different packaging processes, with the Super Extreme Edition emphasizing sustained performance under heavy loads.
In terms of process, both chips are based on TSMC’s 2nm (N2P) node, making them Qualcomm’s first 2nm smartphone chips and delivering the strongest performance seen so far in Qualcomm’s Snapdragon mobile chip lineup.
The most eye-catching figure is the clock speed reaching up to 5GHz.
The Oryon CPU Super Core in the Super Extreme Edition reaches a maximum of 5GHz, while the Performance Core reaches up to 4GHz. Qualcomm states that this is the first mobile CPU to achieve a top clock speed of 5GHz.
However, alongside the clock speed, the FlexCache memory architecture is also getting significant attention this time.
No matter how fast the core runs, if data doesn’t arrive, it can only wait in place.
FlexCache allows different cores to share a dynamically allocated cache pool. When the Super Core encounters heavy tasks, it can access almost the entire cache pool; when tasks are handed off to other cores, relevant data remains in the cache, reducing the need to re-access platform memory.
Agents specifically require frequent handoffs like this. A single travel instruction might involve reading schedules, checking routes, comparing times, and calling services. Different steps are handled by different cores and models in a relay; what FlexCache reduces is the wait time during this handoff process.
GPU changes are also centered around AI.
The newly added Adreno Matrix Cores bring dedicated AI compute cores into the graphics pipeline, working with Adreno Neural Fusion to complete upscaling and frame generation. The 18MB HPM on-chip memory stores intermediate data, reducing the overhead of round-trips to platform memory. This technology is already compatible with Unity and Unreal Engine; Qualcomm also plans to collaborate with 18 mobile games, essentially covering the titles you like to play. In terms of imaging, video specs have been increased to 8K 60fps and 4K 240fps ultra-HD slow motion, with deeper collaboration between the ISP and NPU.
The Hexagon NPU is also shortening the distance between data and compute: shared memory has expanded by 50%, a new Element Accelerator for Transformers has been added, and compatibility with MoE models exceeding 30 billion parameters has been enabled.
As chip performance goes up, thermal management is a major challenge that must be addressed.
Offset PoP moves the packaged memory to one side, creating a dedicated thermal path for the processor area.
From cache to packaging, these changes all address the same issue: smartphones must continue to take photos and play games while also handling new tasks brought by agents.
A clock speed of up to 5GHz raises the compute ceiling, but data access efficiency and thermal efficiency determine how much of that power can be utilized and for how long.
After all, flagship performance can’t just be a "three-minute enthusiasm."
On the PC side, both Microsoft and Google showed up
For personal AI to truly "go to work," it has to start with the computer.
Office work is the first scenario where agents land: reading materials, breaking down goals, opening software, checking results. Tasks are complex enough yet general enough—what’s needed for travel planning and household affairs is actually the same set of capabilities.
So at this year’s Snapdragon Summit, PCs took center stage just like smartphones.
Microsoft brought familiar Surface products: the 12-inch Surface Pro and 13-inch Surface Laptop, powered by the 3nm process Snapdragon X2 Plus, available in Bright Platinum, Roland Purple, and Elegant Black colors. Both models increased screen brightness, and some commercial versions of the Surface Pro offer optional 5G.
The Surface Pro also added Ink Canvas, allowing AI to participate in stylus creation. The Surface mouse revealed at the same time supports haptic feedback and customizable action buttons. These new products are planned to hit the market starting October 13.
Google brought something different: the Googlebook.
Powered by the 4nm process Snapdragon X Elite, it brings the Android ecosystem to laptops and is optimized for Gemini Intelligence. Qualcomm disclosed a battery life of over 19 hours; native apps and cross-device workflows with Android phones are key features of this product line.
One side continues the Windows ecosystem, while the other lands the Android ecosystem. The two paths converge on the same destination: the Snapdragon platform.
Interestingly, Qualcomm also officially announced compatibility with the Linux open-source platform ecosystem. Given the popularity of the AI PC concept, there are many users who choose Linux, and they represent Qualcomm’s potential customer base.
Qualcomm’s goal is to make the 3nm process Snapdragon X2 Elite the key platform defining agent-powered AI PCs—the Snapdragon X2 series delivers up to 85 TOPS of NPU compute, on-device Token generation speed up to twice that of the previous generation, context compatibility expanded to 200,000 Tokens, and plans to demonstrate local execution of models with over 30 billion parameters.
These metrics correspond to different work needs. Long context accommodates materials and task records, while local inference reduces dependence on networks and cloud services. For tasks requiring repeated file reading and continuous content modification, the model must not only compute but also remember.
Office Agents are a competitive focus because office tasks are sufficiently complex yet possess universality.
Reading materials, breaking down goals, operating software, and checking results—these capabilities apply equally to travel planning and household affairs.
The endpoint of the Office Agent competition touches on the capabilities of a General Agent.
Thus, PCs become an unavoidable stop in Qualcomm’s personal AI layout.
Currently, the number of models compatible with Snapdragon PCs has exceeded 450; apps like Claude, Codex, OpenClaw, and Perplexity can already run on the platform, and more on-device AI developers have joined the collaboration.
Surface and Googlebook have combined chips, software, and complete devices. For Qualcomm, running models locally is just the beginning; getting these capabilities into the computers people open every day and handling the work they do every day is what gives Snapdragon PCs the chance to become primary devices for more users.
AI can be imaginative, but office computers must have the ability to deliver.
Earbuds want to take on a bit more, too
After the laptop is closed, life continues.
Checking schedules during commutes, jotting down ideas while walking, asking for advice while shopping—these needs are fragmented, immediate, and often incomplete. For personal AI to participate, it needs to know what is happening in this time and place.
The value of earbuds, glasses, and watches is thus revealed. Sound, first-person perspective, and body status can supplement information beyond the display. Because devices are close to the person, they have the chance to understand background context that isn't explicitly stated.
But wearable devices have a common challenge: tasks can be heavy, but earbuds and glasses cannot be.
The second-generation Snapdragon Audio Platform Extreme Edition replaces the previous S7/S7 Pro. Compared to the previous generation, AI performance is up to twice as high, power consumption is up to 40% lower, and board space is reduced by 30%. For earbuds, the latter two points are just as critical as computing power.
Music, noise cancellation, and calls are already draining battery, and AI needs to stay on standby continuously. The new platform uses a multi-core architecture to support these tasks running in parallel, while integrating ultra-low-power Wi-Fi. Combined with XPAN and Bluetooth to extend connection range, it supports access to cloud intelligence and services.
Earbuds don’t need to complete all computations alone, but they must hear clearly and stay connected.
Qualcomm gave a very specific scenario: wearing earbuds on the street listening to music, noise cancellation should isolate the environment; once it recognizes an approaching siren, the platform adjusts transparency to let key sounds in. The new adaptive ANC also adjusts based on fit, movement, and noise changes.
Noise cancellation can’t just blindly work hard; earbuds need to learn to "listen to sounds and understand events."
Taking it a step further, they can also handle specific tasks.
The "Snapdragon Audio Apps and Agents" ecosystem launched by Qualcomm aims to bring apps and services directly to wearable devices. MicroEJ provides an app framework, allowing capabilities to be added even after models are released; partners like Cinemo are participating in content integration, along with Roon and Qobuz.
Comparing prices, adjusting appointments, and arranging travel are experiences Qualcomm hopes to achieve via voice. Operations that originally required unlocking a screen and finding an app can be replaced by a single phrase.
Sound quality and noise cancellation determine the auditory experience, while perception, connectivity, and apps determine what they can do besides listening to music. Earbuds can take on more tasks, but they shouldn't add extra hassle: speak up when a reminder is needed, stay quiet when things are fine, and hold a charge until you get home.
From earbuds to glasses and watches, for personal AI to enter the gaps in commuting, exercise, and walking, it faces similar challenges. Compute can be hidden in chips, but the sense of presence shouldn't be on your face or hanging from your ears.
In the past, Qualcomm’s impact on smartphones was ultimately reflected in a photo, a video, or a smoothly running game. Consumers may not remember the names of ISPs and GPUs, but they are familiar with the feelings they bring.
Personal AI will follow the same path. Earbuds are responsible for listening, glasses for perception, and phones and computers for thinking and execution—devices have their respective roles but collaborate with each other; AI no longer belongs to a single device, but follows the person.
By that time, what users are buying on the surface is individual pieces of hardware, but in reality, they are choosing a set of intelligence about themselves.
So, the releases for phones, PCs, and earbuds tell the same story: models can come from different vendors, platforms can have different personalities, terminals can carry different brands, but Qualcomm hopes they all have Snapdragon behind them.
From smartphones to PCs, from earbuds to glasses and watches, Qualcomm chips are in every lane. The race for personal AI entry points is still ongoing; no one knows for sure whether the next winner will be the phone, the PC, or the earbuds/glasses.
And what Qualcomm wants to do is ensure that behind all these entry points is the same Snapdragon.