SenseTime launches SenseNova 6.8 Flash, a multimodal agent model, free for a limited time

SenseTime (商汤科技) has released SenseNova 6.8 Flash (日日新), a multimodal agent model that is free for a limited time. The company positions it around "delegated intelligence" (委派智能): the user supplies only a goal, the AI completes the full workflow, and the human handles acceptance. The details below come from a Chinese-language report by QbitAI (量子位), published 2026-11-02.

Benchmark results (SenseTime's figures)

The results below are SenseTime's own. Numeric scores are not given in the report, only rankings.

• First place: TeamBench and DataAnalysis2.0.
• Second place: AutomationBench, WorkbuddyBench, RaccoonPPTBench v2 and BrowseComp.
• Long-term memory: tied for first with GPT 6 Luna on BEAM 128K. It ranks behind Claude 5.5 Haiku on LongMemEval-S and PersonaMem-v2.

Parameter count, context window and pricing after the free period are not yet disclosed.

Hands-on tests

QbitAI ran its tests in SenseTime's Raccoon (小浣熊) client. Setup required installing four plugins (local runtime, browser capability, document preview, video generation) and two skills, raccoon-poster-media and raccoon-dynamic-ppt.

Outputs included:

• A 3D bookstore homepage for a site named "Folio", with books sliding in and reacting to the cursor.
• A brand site for a dexterous robotic hand.
• A horizontal poster for a fictional citrus sparkling water. QbitAI says the animation runs 12 seconds and loops seamlessly.
• A 22-second SVG animation resembling a mine-safety training video, drawn entirely in code.
• An animated launch deck for a fictional telescope, the NOCTIS S80.

QbitAI says these outputs are usable as they are. It also says the model handles webpages that Gemini 4 Carbon struggles with, and notes that some users have criticized that model's web and UI output despite its strong 3D ability. This is QbitAI's assessment, not an independent benchmark.

Memory and proactive interaction demo

QbitAI tested the memory behavior on an internal R&D build of Raccoon. The scenario: a market report and a sales onboarding slide, both made on September 30, cited the same industry survey. The survey said 58% of 1,000 manufacturers use enterprise knowledge bases, so the slide read "more than half".

The agent was then asked only to prepare a chart reprint request for a November 12 client seminar. While checking the research institute's reply, it found that the survey had been corrected on October 30. About 100 trial-only companies without paid contracts had been counted, and the figures changed from 58%/42% to 48%/52%.

The agent linked this to its stored memory and, unprompted, warned that the November onboarding slide still used the old data. After the tester clicked "accept and execute" and asked for the company report to be updated too, it produced two v2 drafts and a revision list, and kept the originals. It adjusted the bar chart proportions and changed the conclusion to "48%, not a majority". Changes remain subject to human approval.

What this means

SenseTime's benchmark selection centers on office automation, multi-step collaboration, data analysis and long-term memory, which matches the delegation framing. The memory demo shows the intended workflow: the model surfaces a stale-data risk from earlier work and leaves the decision to the user. That demo ran on an internal R&D build, so behavior in the public Raccoon client may differ. Rankings are SenseTime's own, with no numeric scores and no independent verification yet. The limited-time free access lets users test the claims directly in the Raccoon client.