A sales rep loses a deal because the product lacked private deployment. A month later the feature ships. Proactive follow-up is the pitch behind SenseNova 6.8 Flash, released by SenseTime (商汤) on 2-3 November — and it lands on the back of the company's first profitable half-year.
Key takeaways
- SenseNova 6.8 Flash adds long-term memory and active interaction, enabling cross-task, cross-session collaboration rather than one-off answers.
- Its delivery forms now span static web pages, dynamic content, and design outputs (Web, SVG, Three.js, Blender, PPT), with Figma workflow support.
- Per IDC's H1 2026 China enterprise MaaS report, SenseTime's large-device private-deployment share is 9.4 percent, ranked TOP 2.
- SenseTime posted a first IFRS half-year profit, reported in RMB in the company's interim results, in H1 2026; generative AI (生成式 AI) made up nearly 80 percent of revenue.
- The company also open-sourced SenseNovaLoop-9B-Preview, a looped-transformer architecture experiment.
From "generate" to "deliver"
This year China's large model (大模型) race has shifted from parameter counts and leaderboard scores to whether an agent (智能体) can actually run a long, multi-step task to completion. Office scenarios — where willingness to pay is clearest and commercialization nearest — have become the contested ground. SenseNova 6.8 Flash bets on three things: that the model remembers, that it acts proactively, and that it can design.
The memory piece is the headline. The model persists facts, preferences, and goals from past interactions; when related information changes, it flags affected items and suggests action. SenseTime's example is a departing employee's handover: based on authorized history, the model drafts which clients are in progress, which needs are unresolved, and what to watch for — continuity that survives staff turnover.
The design piece extends output beyond text and images to structured deliverables. The model can produce Web, SVG, Three.js, Blender, and PPT artifacts, and plug into Figma, aiming at an "AI-native design" workflow rather than a one-shot generation.
The business underneath
The launch is backed by hard numbers. IDC's H1 2026 enterprise MaaS report puts SenseTime's large-device private-deployment share at 9.4 percent, second in the category — and private deployment is widely seen as the main battlefield for China's enterprise model commercialization. SenseTime's H1 2026 results show a first IFRS half-year profit, reported in RMB in the company's interim results, with generative AI (生成式 AI) close to 80 percent of revenue, signaling that growth is now driven chiefly by the new business line. On 5 November the stock rose more than 8 percent intraday, crossing HK$1.2.
The company also open-sourced SenseNovaLoop-9B-Preview, a 9-billion-parameter looped-transformer experiment, a bet on next-generation architectures beyond the dominant recipe.
A profitable generative-AI half-year matters beyond SenseTime's own books. For most of the 2023-2025 period, Chinese model vendors burned cash to win benchmarks and enterprise pilots, leaving the market unsure which players could survive to scale. SenseTime's result — thin by absolute standards, but a signal that the generative line can cover costs — reframes the conversation from "who is smartest" to "who is solvent and deployable." That is precisely the audience sense that matters to CIOs weighing multi-year private-deployment contracts, where vendor longevity sits above any single demo.
According to Li An, chief scientist at BrainNet (脑机网), the shift from score-chasing to task-delivery is the defining move of China's 2026 model race, and vendors that cannot turn model ability into finished work will find their valuations hard to defend.
The result also resets expectations for the wider field. Peers pursuing the same enterprise-private-deployment wedge — from the cloud units of Baidu and Alibaba to dedicated model labs — face similar pressure to show that pilot revenue converts into durable margin rather than one-off project fees. SenseTime's numbers do not prove the category is solved, but they remove one argument against it: that generative AI at enterprise scale is permanently loss-making. For CIOs, the takeaway is less "buy SenseTime" than "the private-deployment bet is now demonstrably survivable, so diligence should shift to integration depth and data-governance posture, not existential questions about vendor survival."
What readers can do now
- If you run knowledge management or sales handover, trial 6.8 Flash's long-term memory — but define the memory boundary and privacy authorization first.
- Judge vendors by private-deployment share and revenue mix, not just benchmarks. SenseTime's 9.4 percent and ~80 percent generative-AI revenue are more telling than a single score.
- Treat the open-sourced Loop-9B-Preview as research, not a drop-in commercial flagship.
Honest limitations
This article draws on SenseTime's official site, the Economic Daily's China Economic Net, and Oriental Network reporting. The first IFRS half-year profit, 9.4 percent private-deployment share, near-80 percent generative-AI revenue, and the >8 percent rise to HK$1.2 are reported as those sources state them; the profit is reported in RMB in the interim results, and I have not made an independent currency conversion. Long-term-memory and active-interaction claims are demonstrated by SenseTime and not independently benchmarked here; industry observers note that memory boundaries and privacy protection remain open questions at deployment. I have not verified SenseTime's specific benchmark comparison objects or the exact license terms of the open-sourced model.