SenseTime Releases Multimodal Agent Model SenseNova 6.8 Flash

Truly, being genuinely useful is what really matters.

And the limited-time free access isn't even the main point—the key is that it can completely handle web pages that even something as powerful as Gemini 4 Carbon can fail at!

For example, a website introducing cameras can be complete and stylish to this extent:

After all, quite a few people have been complaining that although Gemini 4 Carbon is strong at 3D, its web pages and UI... Emmm...

And this new domestic AI can handle not only web pages, but even the PPTs we commonly use for work—and in 3D dynamic form, no less~

And also dynamic posters:

Of course, it can also handle a lot of purely 3D content now:

To be fair, whether it's websites, dynamic PPTs, or dynamic posters, they can all be said to be at a level that can be used directly.

Without keeping you in suspense, it is SenseNova 6.8 Flash, the multimodal agent model just released by SenseTime, whose main focus is delegated intelligence—that is, users only give the goal, the AI goes through the entire process itself, and humans are only responsible for acceptance.

As for how strong it really is, let's first look at a round of official benchmark results.

In benchmarks such as office automation, multi-step collaboration, and data analysis, SenseNova 6.8 Flash took first place on TeamBench and DataAnalysis2.0, and also ranked second on AutomationBench, WorkbuddyBench, RaccoonPPTBench v2, and BrowseComp:

And in evaluations testing long-term memory, it tied for first with GPT 6 Luna on BEAM 128K, and was second only to Claude 5.5 Haiku on LongMemEval-S and PersonaMem-v2:

Let me emphasize one key point again: it is currently limited-time free!

So how does SenseNova 6.8 Flash actually feel in practice?

Come on, let's dive into a round of hands-on testing~

As usual, before we begin, let's briefly introduce the operating environment and method.

This hands-on test of SenseNova 6.8 Flash was all completed in SenseTime's Xiaohuanxiong client, and we first need to complete a few simple settings.

First, in the "Settings" > "Plugins" section, install these four components: local runtime environment, browser capability, document preview component, and video generation:

Then go to the "Capability Center," find "Skills," and install raccoon-poster-media and raccoon-dynamic-ppt below:

After these simple settings, we can create a new task and start the hands-on test.

The first hands-on test we want to do is dynamic web design.

Just in the past couple of days, X blogger @viktoroddy posted a dynamic bookstore official website made with AI, so we directly asked SenseNova 6.8 Flash to replicate one as well.

The prompt was roughly: "A 3D bookstore homepage hero screen with a row of gradually rising books as the visual subject, with the official website name Folio."

Let's take a look at the result:

As you can see, when we open the page, the large word "Folio" slowly becomes clear on a black background, and then a row of books slides in diagonally from the lower left corner, like a row of dominoes half fallen.

And when you move the mouse over a book, that book pops outward—it is indeed an interesting bookstore official website.

Next, let's make a dynamic brand official website.

This time, we changed the subject from books to a dexterous hand for embodied intelligence.

In the middle of the screen, a black bionic hand slowly rises from the bottom, its fingers gradually opening and closing, and the metal joints reflect light as they move.

In the lower left corner there is just one English tagline and a booking button, with everything else left blank. To be honest, if you mixed it in with a bunch of real brand official websites, it would be hard to tell it apart.

After web pages, let's look at the effect of dynamic posters.

We asked SenseNova 6.8 Flash to make a horizontal product poster for a fictional citrus sparkling water:

The moving parts in the poster are all hidden behind the bottle: an orange circle slowly spreads out, and a few orange slices rotate in from both sides and then out again. The whole segment is 12 seconds, with the beginning and end connecting perfectly, so it can be looped directly.

In addition to web pages and posters, SenseNova 6.8 Flash can also make SVG animations.

Let's take a look at the result:

The whole segment is 22 seconds with no dialogue at all, but the story is told clearly and completely, and it looks like a mine safety training animation.

Moreover, the swinging of the character's hands and feet while walking and the shift in center of gravity when squatting are quite natural, and the camera also pushes in and pulls out to follow him. You should know that all these images are vector graphics drawn stroke by stroke entirely with code.

Finally, let's do a hands-on test of the very practical dynamic PPT for work.

We asked it to make a launch event PPT for a fictional telescope, with the product called NOCTIS S80:

With results like this, if we use it for proposals or internal reviews, the level of completion of SenseNova 6.8 Flash is already fully sufficient.

Of course, design is one thing—but what other, more complex work can SenseNova 6.8 Flash do for working people?

For this part, we tested it in advance in SenseTime's Xiaohuanxiong (internal R&D version), and found that SenseNova 6.8 Flash also has a knack for intelligent office work.

And what it emphasizes most is long-term memory and proactive interaction.

The general background of the task was like this:

A month ago, that is, on September 30, the company produced a market analysis report and a one-page sales onboarding training material, both of which cited data from the same industry survey: among 1,000 manufacturing enterprises, 58% were already using an enterprise knowledge base. So the training page said "more than half," and this page of material will continue to be used in November.

Xiaohuanxiong remembered all this previously completed work in its memory.

Then, now, we asked it to do another thing: prepare a chart reproduction request for the client seminar on November 12.

Let's look at the result:

As you can see, when the Agent checked the research institution's reply, it found that the other party had corrected this survey on October 30. It turned out that 100 enterprises that were only trialing and had not signed paid contracts had been included, and the data was changed from 58%/42% to 48%/52%.

You should know that the task we gave it this time was only a reproduction request; we did not ask it to check the training materials.

But it connected this correction with the memory from a month ago and proactively popped up a reminder: the sales page to be used for November onboarding training was still using old data, and it suggested checking and updating it first.

We clicked "Accept and Execute" and also asked it to revise the company report as well.

Next, it generated two v2 drafts for review and a revision list, and the originals were all retained as well. Moreover, it was not just the numbers that were changed; the proportions of the bar chart were adjusted accordingly, and the conclusion "more than half" was also changed in sync to "accounting for 48%, not more than half."

Situations like this, where the data has been updated but no one remembers to revise the old materials, are actually quite common in companies.

What SenseNova 6.8 Flash does is remember these things on people's behalf and remind them at the appropriate time. As for whether to make the changes, it is still up to people to decide.

For example, just today, GPT-6.2 Astra opened gray testing again, and Claude Mythos 5.5 was also exposed. The pace of new model iteration really is getting faster and faster.

But however strong they are, a very realistic problem is that many people cannot experience them right away.

Just like Gemini 4 Carbon, which Google has already released for some time: so far, only a small number of people can use it, and most people can still only watch others show off their results on X.

And with the webpages, posters, and PPTs we made using SenseNova 6.8 Flash, you can now open Xiaohuanxiong and try them out yourself.

And in design, what we can feel is that the barrier has been lowered. Previously, if you wanted to make an official website with 3D motion effects, you needed a designer to produce visuals and a front-end developer to write code, going back and forth for several rounds of revisions; now you just need to explain the desired effect clearly and leave the rest to AI.

And these two things ultimately point to the same word: delivery.

This change can actually also be felt recently from the releases of domestic large models: there is less pure showing off of benchmark scores, and more and more directly throwing out a finished webpage or a revised document.

In a word, the upper limit of a model is defined by the big companies; the value of a model is defined by the people who can actually use it.