SenseTime launches SenseNova 6.8 Flash, a multimodal agent model that is free for now

SenseTime has released SenseNova 6.8 Flash, a new multimodal agent model that is free to use for a limited time. Rather than leading with parameter counts or benchmark tables, the company is pitching “delegated intelligence”: the user states a goal, the model plans and carries out the entire workflow, and the person only signs off on the result.

SenseTime’s published benchmark results put SenseNova 6.8 Flash first on TeamBench and DataAnalysis2.0, and second on AutomationBench, WorkbuddyBench, RaccoonPPTBench v2 and BrowseComp. On long-context memory it ties GPT 6 Luna for first on BEAM 128K and trails only Claude 5.5 Haiku on LongMemEval-S and PersonaMem-v2.

The model is accessible through SenseTime’s Raccoon client. Users first install four plug-ins — a local runtime, browser capability, document preview and video generation — then add the raccoon-poster-media and raccoon-dynamic-ppt skills from the capability center before creating a task.

In demos, a single prompt produced a 3D bookstore homepage called Folio, with the title fading in over a black background and a row of books sliding in diagonally; hovering over a book makes it pop outward. The model also generated a brand site, a 12-second looping product poster, a 22-second SVG animation and a dynamic slide deck.

The office scenario is the more consequential test. QbitAI described a cross-month task in which a market report and a sales onboarding page both cited a survey finding that 58% of manufacturers already used an enterprise knowledge base. The research firm corrected the figure to 48% on October 30. Asked only to prepare a chart-reprint request, the agent connected the correction to its earlier memory, warned that the November training material still used the old number, and, after approval, produced two revised drafts plus a change list.

That makes long-term memory and proactive interaction feel like product features rather than benchmark line items. Work once split between designers, front-end developers and several review rounds can now collapse into two steps: describe the requirement, then approve the output.

For SenseTime, the pitch is less about topping another leaderboard than about handing ordinary users an agent that can finish and deliver a whole task. The open questions are how long the free window lasts, how stable the model is on messy real-world tasks, and whether it holds up under repeated daily use.