Author | Wang Han Editor | Yun Peng
ZhiDongXi reported on November 3 that the official release of SenseTime's multimodal agent model, SenseNova 6.8 Flash, was announced yesterday evening. Leveraging long-term memory and proactive interaction capabilities, this model demonstrates significant advantages in agent collaboration, office intelligence, and design.
According to disclosures by the SenseTime team, SenseNova 6.8 Flash outperforms models such as Claude Opus 5, GPT-6 Sol, and Kimi K3 in benchmark tests evaluating team collaboration, data understanding, analytical reasoning, and result delivery.
Following the model's release, ZhiDongXi immediately conducted a hands-on test of SenseNova 6.8 Flash within the desktop version of SenseTime's AI office agent, Xiao Huanxiong (Little Raccoon), focusing on office scenarios.
Our testing revealed that the model performs well in capabilities such as long-context retrieval and visualization. Additionally, the visual language of its generated PPTs and posters remains consistent.
In terms of generation time, SenseNova 6.8 Flash spent considerable time repeatedly checking during the latter half of the PPT generation process, resulting in extended duration. This may be attributed to environmental and harness influences; therefore, this result does not directly represent the model's inherent generation speed.
01.Going Head-to-Head with GPT-6 Sol and Claude 5.5 HaikuRanked First in Team Collaboration and Data Analysis
In benchmark testing, SenseNova 6.8 Flash ranked first among the listed models in TeamBench, which tests team collaboration capabilities, and DataAnalysis2.0, which evaluates data understanding, analytical reasoning, and result delivery, surpassing models such as Opus 5, GPT-6 Sol, Kimi K3, and GLM-5.3 Flash.
In four other tests—AutomationBench for automated task execution, WorkbuddyBench for work assistant collaboration, RaccoonPPTBench v2 for content organization and visual expression in PPT generation, and BrowseComp for web browsing and retrieval—SenseNova 6.8 Flash maintained second place, trailing only behind GPT-6.5 Sol while outperforming Gemini 3.1 Pro, Claude Opus 5, and others.
In LongMemEval-S, which tests long-text memory capability, SenseNova 6.8 Flash ranked second with a score of 93.33, surpassing GPT-6 Luna and DeepSeek V4.1 Flash.
In BEAM 128K, which tests long-context understanding, SenseNova 6.8 Flash tied for first place with GPT-6 Luna; in PersonaMem-v2, which tests role memory and personalized interaction, it ranked second, trailing only Claude Haiku 5.5; while in π-bench, a specific comprehensive reasoning/planning test, it ranked fourth with a score of 58.16.
02.Reads Two Prospectuses in 8 MinutesAnd Creates Interactive Dashboards
With such impressive benchmark scores, how does it perform in actual use?
We tested SenseNova 6.8 Flash on the SenseTime Xiao Huanxiong desktop client, focusing on office scenarios.
Recently, Anthropic released its IPO prospectus and updated the latest data yesterday. Combined, these two prospectuses total 582 pages, with much of the data buried within the text, making manual extraction tedious.
We tasked SenseNova 6.8 Flash with comparing the two prospectuses to identify updated data points and generate a data dashboard. This task rigorously tests the model's context retrieval, information extraction and entity alignment, structured output and data visualization capabilities, and crucially, financial knowledge. It can be observed that SenseNova 6.8 Flash took a total of 8 minutes and 21 seconds to complete these two tasks.
The model identified that the latest prospectus contained six key data updates: interim revenue, interim net loss, cash and short-term investments, total assets, accumulated deficit, and Google compute commitment amount, providing before-and-after comparisons for each. Upon verification, all figures listed by SenseNova 6.8 Flash were correct.
The data dashboard generated by SenseNova 6.8 Flash was also comprehensive, consisting of five sections: Update Overview, Financial Performance, Assets and Liabilities, Operational Metrics, and Commitments and Updates. Each page included corresponding before-and-after data comparisons, with bar charts and line graphs created for key financial metrics. Hovering the cursor over chart elements displays the specific data point values.
PPT creation is a common challenge in office scenarios. We asked SenseNova 6.8 Flash to create a PPT highlighting key data from the prospectus, using the color scheme from the prospectus images. For this task, the model worked for 64 minutes. Reviewing SenseNova 6.8 Flash's thought process, we found that the majority of the time was spent on final formatting adjustments, repeatedly tweaking chart sizes, page margins, and font sizes. Of course, the extended duration may also be influenced by the environment and harness, and does not directly reflect the model's intrinsic output speed.
Returning to the task, upon receiving instructions, SenseNova 6.8 Flash first defined an 8-page structure and outlined the basic narrative flow. The color palette was sampled from the prospectus images, adopting a low-saturation editorial style with warm beige backgrounds, pink, mint green, light blue, and warm gray.
03.Supports Dynamic Posters and Agentic 3D DesignAlso Integrates with Figma
According to the SenseTime team, SenseNova 6.8 Flash's design generation supports formats including Web, SVG, Three.js, Blender, and PPT, and also enables dynamic PPTs, dynamic posters, and Agentic 3D design. For example, when creating PPTs, SenseNova 6.8 Flash first analyzes the information to be conveyed on each page, where pauses should occur, and what needs emphasis, then orchestrates animations accordingly.
For dynamic posters, SenseNova 6.8 Flash can holistically consider main visuals, text hierarchy, and animation orchestration within a single generation pass: the motion amplitude of the main visual coordinates with the position and timing of text appearances, with timeline rhythms serving the information hierarchy.
In 3D modeling, SenseNova 6.8 Flash can complete complex 3D creations through Agent coordination with Blender. After users input requirements, the Agent generates an initial scene version and continuously adjusts based on actual rendering results and user feedback. Users can request specific scenes within the Blender software, such as the Louvre, and then correct flaws in the initial output, such as floating roofs or pyramid positioning. Users can also add modification requests, such as weather animation effects. Through multiple rounds of "generation—observation—feedback—modification," dynamic 3D scenes are progressively refined, ultimately outputting an editable Blender project file.
In SVG design, the model also supports vector graphics and animation creation, combining shape construction, color matching, path control, and animation orchestration to generate stylistically consistent, losslessly scalable, and secondary-editable vector designs.
Simultaneously, SenseNova 6.8 Flash supports 3D scene and interactive content creation, comprehensively processing spatial layout, materials and lighting, camera control, and animation orchestration to build 3D content that renders in real-time on the web.
Beyond workstation design, the SenseTime team noted that the model can integrate with workflows like Figma, helping users embed AI design capabilities into existing professional design processes.
04.Conclusion: Models Now Possess Delivery CapabilitiesMay Accelerate Office Adoption
Based on our testing results, SenseNova 6.8 Flash demonstrates high usability in long-text processing, complex information extraction, and multimodal design. Although tasks such as complex PPT layout take longer, its ability to directly output editable files indicates it already possesses delivery capabilities.
Just days ago, SenseTime also open-sourced SenseNova Loop-9B-Preview, providing an open-source implementation of a 9B-parameter Looped Transformer.
Combining its exploration of frontier architectures and investment in the open-source community, SenseTime appears to be accelerating the deployment of AI in real-world office and professional design workflows through a dual-track strategy: "closed-source models targeting high-complexity applications, and open-source models exploring next-generation architectures."