Generative AI already supplies nearly 80% of SenseTime's revenue, and 6.8 Flash pushes that business toward paid office agents once free use ends on December 1.
SenseTime has released the official version of SenseNova 6.8 Flash, a multimodal agent model that it is pitching at office collaboration and design work. The company announced it on the evening of November 2, IT Home reported. The model adds long-term memory that carries across sessions and tasks, and it can turn a plain-language request into an animated PowerPoint deck, a motion poster, a web page, an SVG animation or a 3D scene built in Blender. It can also connect to design tools such as Figma.
The model is free to use in Raccoon, SenseTime’s AI assistant app, until December 1, and is also available through the SenseNova Token Plan, SenseTime’s subscription for developers and businesses. SenseTime did not publish standard pricing for the period after the free trial.
SenseTime also gave no benchmark scores. It said only that on several benchmark evaluations the model beat some international frontier models on agent collaboration, data analysis and presentation generation, without naming the benchmarks or the models or giving figures, The Paper reported. Readers have the company’s demonstrations to go on, not independent results.
SenseTime pitches the model at finished office deliverables
The release is built around a working method SenseTime calls "delegated intelligence": the user sets a goal, the agent calls the tools it needs and produces the deliverable, and the user reviews the result, according to TechNode. TechNode’s demonstrations in the Raccoon client included a bookstore homepage with 3D motion effects, a brand site for a robotic dexterous hand, a looping drinks poster, an SVG animation written in code and an animated deck for a product launch.
The slide feature plans a deck before animating it. IT Home said the model first works out each slide’s key message, pauses and points of emphasis, then generates transitions, the order in which elements appear, chart builds and layout together, and exports the result as a PPTX file. Posters are produced the same way, with the main image, text hierarchy and motion generated as one piece and a visual style that can be extended into a series.
For 3D work, an agent operates Blender, produces a first scene and then revises it over several rounds based on the rendered images and the user’s comments. The Paper said the output is an editable project file, but the feature requires a local software setup and several rounds of interaction.
Memory is meant to keep long projects from losing track
The other headline feature targets a common failure of AI assistants on multi-stage work: forgetting earlier decisions. SenseNova 6.8 Flash keeps project requirements and past information, links new information to needs the user recorded earlier, and suggests changes when something relevant shifts.
SenseTime gave two examples. In one, described by The Paper, the model notices after a product upgrade that some customers had earlier declined to cooperate because a feature was missing, and prompts staff to contact them again. In a media test reported by TechNode, the agent was asked to reuse a chart, found that a research institute had since corrected its survey data, and flagged earlier reports and training materials that still used the old figures. With the user’s approval, it drafted revisions and a list of changes covering figures, chart scales and conclusions, while keeping the original files.
TechNode noted that the memory and proactive-collaboration tests ran on an internal research version, while the web, poster and slide demonstrations ran in the public Raccoon client.
SenseTime already earns most of its revenue from generative AI
The company has tied its recovery to generative AI. SenseTime’s generative AI revenue reached RMB2.33 billion in the first half of 2026, or 79.9% of total revenue of RMB2.91 billion. The company reported a profit of RMB617.3 million for the period, its first since listing, helped by fair-value gains on companies it incubated or invested in, while its adjusted net loss narrowed 67.3% to RMB385.9 million, TNGlobal reported. EastFrontier covered how SenseTime turned its AI bet into a first-half profit.
In its results announcement, SenseTime said average daily token volume on its SenseCore platform reached 2.4 trillion in July, up about 22-fold year on year. The Token Plan that now carries 6.8 Flash opened in May alongside SenseNova 6.7 Flash-Lite, which SenseTime said cut token use by up to 60% in information-search tasks compared with text-only agents, according to its 6.7 Flash-Lite announcement.
The price SenseTime sets for 6.8 Flash once the free Raccoon period ends on December 1 will be the clearest sign of how it expects office agents to contribute to revenue.