新智元报道
最近，大模型圈的风向，彻底变了。
以前，大家还在比拼模型能生成多少内容、手搓多少行代码；然而，现在的用户，早就不想要只停留在对话框里的「半成品」了。
大家真正看重的是：它究竟能不能帮我们干完一件具体的工作任务？能不能听懂复杂需求、持续推进任务，并最终交付真正拿来即用的成果？
好在，就在这几天，商汤祭出大招，推出SenseNova 6.8 Flash正式版！
新模型已全面接入商汤小浣熊桌面客户端与Token Plan。
而且，一个Flash级别的轻量级模型，居然在办公协同与视觉工程交付基准上，对标国际顶尖模型！
目前商汤小浣熊客户端可支持体验3D，Web，动态Poster等功能。
在多项权威基准测试中，SenseNova 6.8 Flash的表现非常抢眼。
多Agent协作（TeamBench）、数据分析（DataAnalysis 2.0）、办公协同（WorkbuddyBench）、深度检索与演示文稿（BrowseComp & RaccoonPPTBench v2） 等多项能力， 对标Opus 5.5与GPT 5.6 Sol等海外顶尖模型。
跑分确实惊艳，但如果把复杂的设计和办公需求交给它，它能否交出带审美的成品？
接下来，我们打开小浣熊客户端，从矢量物理联动、多模态空间寻路，到动态镜头调度与数据叙事，用几道变态级考题，开展了真刀真枪的实战。
视觉生成的第一关，我们直接跳过简单的静态排版。
我们要测的是：模型能不能用纯代码算准机械物理联动？能不能看懂复杂插画里的三维结构？
「用SVG代码画一只骑自行车的鹈鹕」，这可是AI圈的保留题目。
真正的难点根本不是画一只鸟，而是机械结构的几何对位与运动学联动。
很多AI做出来的demo，车轮在转但鸟腿悬空，非常诡异。
这次我们加大难度，不仅要求小浣熊用内联SVG手绘三只一前两后、鹈鹕，而且定下严苛的物理约束——
曲柄、脚踏与双段式鸟腿必须铰接联动；抬车头时，整车必须绕后轮轴心旋转。
经过多轮优化后，效果十分惊艳。
画面中，三只鹈鹕的膝关节随着脚踏的旋转，完成了精准屈伸。按下「冲一段」、「下点雨」和「抬前轮」时，也是相当自然。
能把连杆动作算得如此丝滑，可见模型用纯代码构建矢量图层、做空间编程的功底相当扎实！
这道题的关键，是看懂插画遮挡，自动规划矢量寻路。
我们传入一张暖光微距的「立体纸艺书店插画」，要求小浣熊识别出画中6个书摊的方位与中间走道，并写出一套「纸飞机寻路系统」，同时实现昼夜灯光切换、摊位现场改名以及SVG矢量路线图导出功能。
其中，一个关键要求是：纸飞机飞向摊位时，需要沿着真实过道拐弯，避开墙体遮挡。
当我们在页面点击特定摊位时，一只白色纸飞机立刻顺着院落走道，平滑绕开桌角，稳稳停在目标上方。
切到「夜灯」模式，橘色窗光精准地落在了六座纸屋的窗框内。
要知道，喂进去的只是一张静态插画，它却能一眼看穿画里的走廊动线与建筑遮挡，这种多模态空间感知力简直绝了！
这一关的题，重点考察的是模型能不能用纯代码搭建3D光影空间，并做到多机位镜头调度与跨模态的数据耦合。
首先是，不给任何3D模型素材，只靠纯代码，让小浣熊手搓一个带多机位镜头的雨夜3D街景！
模型生成的3D光影质感与镜头感，直接拉满——
点击「进站开门」，墨绿色复古电车亮着黄色大灯穿破雨雾，平稳滑入站台。
再点击「关门离站」，车门闭合，电车缓缓提速驶入夜色，空间景深与时间节奏一气呵成，十分自然。
虽然仔细看会发现，电车驶离时和路边的电线杆还有点小穿模；但不依赖外部建模软件、仅凭纯代码就能生成这种雨夜运镜，整体完成度已经很高了！
考题4：暗场后台彩排展演
这个考题中，我们上传了一张工业风舞台后台底图，附上8位创作者展品资料与排期表。
要求模型定位书本、植物小灯等8件实物坐标，写出让舞台光束、资料卡与时间轴三方联动的聚光灯遮罩系统。还要外加一个能沿虚线撕开的入场票。
模型给出了一个空间分布颇为精确的演出后台。
按下「顺序预演」，聚光灯会精准锁定道具依次扫台，时间轴滑块同步跳转。
点开票夹，真的能亲手撕下副券并导出高清票面。
一套操作下来，SenseNova 6.8 Flash端到端的全栈工程交付能力，得到了直观展现。
接下来，我们考察的是，AI能否把数字变成有交互、有故事的桌面小工具。
这道题中，我们把8笔订阅扣费清单和一张复古公寓收据机底图交给模型。
要求是把8项订阅映射进机身的8扇窗户。点窗户会切换续订状态，触发打印头就会扫过动画并实时重算月费。
实测出来的效果很生动。
点击关掉3号和8号窗户（暂停订阅），热敏打印头会立即扫过，伴随动画吐出新纸条，精确算出一年省下多少月费。
把单次任务做成高审美成品，只是一部分。
SenseNova 6.8 Flash的B面，则是超强的跨周期长期记忆与主动交互能力。
据官方权威基准评测：
128K超长上下文记忆（BEAM 128K）、长程记忆（LongMemEval-S）、主动交互（π-bench）等能力位居行业头部。
换句话说，你的AI终于会「自己盯进度」了！
接下来，我们来看一个真实办公案例。
11月5日，公司要召开重要的新品发布会，市场部和客户对接团队早已基于当前演示环境发出了邀请函、备好了演示账号。这些历史安排，小浣熊已经默默记在了本地记忆库里。
突然，你收到一封运维邮件：「原有用于用户体验的旧环境将在11月3日提前回收。」
然而，其他部门还不知道这件事！如果市场部门和客户对接团队在3号之后还用旧环境给客户做演示，绝对会大翻车。
我们通过商汤小浣熊（内部研发版）进行了测试，去邮箱整理最新通知，指令里半个字都没提发布会。
好在模型拉起浏览器读取邮件后，自动关联历史记忆，一眼揪出「停服比发布会早2天」的冲突，立马在桌面右上角主动弹窗发出预警！
在你点击「接受并执行」确认后，小浣熊直接在后台接管浏览器，一键生成涵盖账号、数据与入口迁移的紧急通知和应对方案，精准填入相关团队收件人并保存至草稿箱待审，把一场危机消灭在了萌芽状态！
一款Flash级别的轻量模型，为什么能同时兼顾前端动效、镜头语言与长程记忆？
答案就藏在商汤的底层架构创新与Agent工程生态中。
几天前商汤开源的SenseNova Loop-9B-Preview展示了一条新路线：「用计算深度换取参数效率」。
这颗模型总参数量约90亿，拥有32层独立解码器层。其核心秘诀就在于——中间的24层被设计为共享权重的循环计算块。
单次前向传播时，这24层会连续循环执行2次，使实际计算深度达到56层，推理等效参数规模一举拉升至14.6B！
配合混合注意力设计，原生支持256K上下文窗口。以九十亿的体量，越级匹敌三倍体量的开源模型。
而在应用生态层，商汤在GitHub开源了SenseNova-Skills技能仓库。
把信息图生成、多风格动态PPT、Excel大表工作流打包成十项标准模块，直接内置进小浣熊客户端。
这就是为什么它做出来的东西这么好看、这么好用。
大模型的下半场，谁能听懂复杂需求，把「视觉美感」与「工程数据」一次性做成完整作品，谁就拿到了入场券。
在这条赛道上，SenseNova 6.8 Flash已经交出了一份颇具诚意的答卷。
目前，SenseNova 6.8 Flash正式版已上线SenseNova Token Plan，设计功能也已经在小浣熊桌面客户端限时免费体验。
属于打工人和设计师的时代已经来了，让SenseNova 6.8 Flash成为你趁手的工作搭子吧！
编辑：Aeneas 摩西
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Recently, the wind direction of the LLM circle has completely changed.
In the past, you've been generating more content and hand-stamping more line codes than a puzzle model; however, users today have longed not to remain in the "half-finished" in the dialogue box.
What is really important is: can it do us a specific job? Can we understand complex demands, sustain mandates, and ultimately deliver what is really being achieved?
Good thing, it's just a couple of days since the soup festival launched Sensenova 6.8 Flash!
The new model is fully accessible to the raccoon desktop client with Token Plan.
Moreover, a flash-level lightweight model can match the international top-notch model in office synergy and visual engineering delivery benchmarks!
The business raccoon client currently supports the 3D, Web, Dynamic Foster, etc.
In many authoritative benchmark tests, Sensenova 6.8 Flash performed in an impressive manner.
Many Agent collaborations (TeamBench), data analysis (DataAnalysis 2.0), office synergies (WorkbuddyBench), deep search and presentation (Browncomp & RaccoonPPTBench v2) have many capabilities, such as the top overseas models of Opus 5.5 and GPS 5.6 Sol.
It was amazing to have run points, but if it was given complex design and office requirements, could it hand over the finished product with aesthetics?
We then opened up the raccoon client, from vector physics to multi-modular space to dynamic shot movement and data narratives, to engage in real-life guns with several perverted tests.
The first level of visual generation, we skip a simple static layout.
So what we're trying to measure is, can the model compute mechanical physics with a pure code? Can you understand the three-dimensional structure of the complex illustration?
"A cycling pelican with SVG code" is a reserved title for the AI Circle.
The real difficulty is not to draw a bird at all, but the geometry of the mechanical structure is associated with motorism.
A lot of AI made demo, the wheels were spinning but the legs were hanging, very weird.
This time we have made it more difficult not only to ask the raccoon to draw three of the first two pelicans with the inside of the SVG, but also to impose severe physical constraints.
The handle, foot and double legs of the bird shall be connected and the whole vehicle shall rotate around the rear axis when carrying its head.
After many rounds of optimization, the effects were spectacular.
In the image, the knee joints of three pelicans were rotated with the foot, and the precision stretch was completed. It's also quite natural to press "drive" and "drop" and "lift the wheel."
It's so smooth that the model can build a vector layer with a pure code and have a solid foundation for space programming!
The key to this is to read the illustrations and automatically plan the vector path.
We have introduced a warm-lighted "Step Paper Books Adventure" that requires the raccoon to identify the location and middle lanes of the six bookstores in the picture and to write a "paper plane search system" with day and night light switching, renaming of the booth site and export of the SVG vector road map.
One of the key requirements is that when paper planes fly to the booth, they turn around along the real aisle and avoid wall cover.
When we click on a specific booth on the page, a white paper plane immediately follows the courtyard walkway, smoothly bypassing the corner of the table and steadying itself above the target.
In the "nightlight" mode, orange window light landed precisely in the window frame of six paper houses.
You know, it's just a static illustration, but it can see through the corridor artery and the architecture.
This issue focused on the ability of the model to build 3D photospace using pure codes and to align multi-space shot movements with cross-modular data.
First of all, without giving any 3D model asset, just using pure codes, let the raccoon's hand rub a 3D view of rain night with multiple planes shot!
The model produces a 3D photometric sense and a shot sense, and it's full directly.
Click "Enter the station and open the door" and the ink-green retrobus is wearing a big yellow light through the rain fog, smoothing down the platform.
It's natural to click "Close the station" and close the doors, and the trams are moving slowly into the night, with a deep space and a rhythm of time.
Although a closer look reveals that the electric train leaves and the utility poles on the side of the road are still a little smudged; without relying on external modelling software, which can be generated only by pure code, the overall completion is already high!
Question 4: Show in the back of the dark field
In this examination, we uploaded a bottom map of the industrial wind stage with eight creators' exhibits and schedules.
Eight physical coordinates, such as model locator books, plant lamps, etc., are required to write a spotlight mask system that connects the stage beam, information card and time axis. And a ticket that can be ripped off along a dead line.
The model gives a fairly accurate spatial distribution of the backstage.
By "sequencing preview", the spotlight locks the props in a sequential sweep, and the time-axis slide jumps in parallel.
You can really rip the coupons off with your own hands and export the high-resolution papers.
A set of operations, with Sensenova 6.8 Flash end-to-end all-storey delivery capabilities, was visualized.
Next, we looked at whether AI could turn numbers into interactive, story-telling little desktop tools.
On this subject, we gave the model eight subscription deduction lists and a copy of the bottom of the receipt machine.
The request is to map eight items into eight windows of the fuselage. The window will be converted, the head trigger will sweep and the monthly fee will be recalculated in real time.
The results are vivid.
Click to close windows 3 and 8 (suspension of subscriptions), and the hot-sensitized printhead will be cleaned immediately, accompanied by an animation of new notes, with an accurate calculation of how much savings are made during the year.
Making a single mission a high-quality aesthetic, just part of it.
The B side of Sensenova 6.8 Flash is an ultra-cyclical long-term memory and active interactive capability.
According to official authority benchmarks:
context (BEAM 128K), Long-Range Memory (LongMemEval-S), Active Interactivity (π-Bench) are at the top of the industry.
In other words, your AI. will finally keep an eye on progress!
Next, let's look at a real office case.
On 5 November, the company held an important new publish meeting, and the Ministry of Markets and the customer matching team had already issued invitations and prepared demonstration accounts based on the current demonstration environment. These historical arrangements, the raccoons have been recorded in local memory banks.
All of a sudden, you get a handy mail: "The old environment used for user experience will be recovered early on November 3rd."
However, other departments have no idea! If the market sector and the customer docking team were to demonstrate to the customer after the third, the old environment would definitely turn over.
We ran a test through the merchant raccoon, and we went to the mailbox to sort out the latest notice, and there was no mention of publish in the instructions.
It's good that after the model pulls up the browser to read the mail, it automatically connects to the historical memory, pulls out a conflict that "suspends two days before publish" and alerts the active window at the top right corner of the desktop!
After you click "Accept and Execute" confirmation, the raccoon takes over the browser directly from the backstage, generating an emergency notification and response plan covering account numbers, data and access movements, accurately filling in the relevant team recipients and storing them in the draft box, and eliminating a crisis in its infancy!
A flash-level lightweight model, why combine front-end kinetic, shot language with long-range memory?
The answer lies in the bottom-up structural innovation of merchant soup and in the eco-engineering of Agent.
A few days ago, SenseNova Loop-9B-Preview, who had the open source of soup, demonstrated a new route: "The depth of calculation for parameter efficiency".
The model totals about 9 billion parameter count and has 32 layers of independent decoder. The core secret is that the center is designed as a circular calculator of shared weights.
When a single forward transmission is made, the 24th floor will be recycled twice, bringing the actual computation depth to 56 layers and the magnitude of the reasoning equation to 14.6B!
The original supports 256 Kcontext window as part of the hybrid design. An open source model of three times the size of the superiors in 9 billion.
At the applied ecological level, commercial soup is the source of the Sensenova-Skills skill warehouse in GitHub.
A multi-style dynamic PPT, Excel large table workflow is packaged into 10 standard modules and placed directly into the raccoon client.
That's why it makes things so nice and so good.
In the second half of LLM, anyone who understands complex demands and makes one-time complete works of "visual beauty" and "engineering data" gets tickets.
On this track, Sensenova 6.8 Flash has submitted a reply in good faith.
Currently, Sensenova 6.8 Flash, the official version of Sensenova Token Plan, is on line and the design functionality is already being experienced free of charge by limited-time, the desktop client of the raccoon.
The time has come to beat up workers and designers, and let Sensenova 6.8 Flash be your working partner!
Editor: Aeneas Moses
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.