A shopper snaps a shelf photo; a model reads the prices, computes the best per-unit deal across apps, and says the answer out loud in real time. The demo is mundane on purpose. SenseTime is betting that useful media intelligence looks less like a magic clip and more like a tool that quietly understands images, sound, and text together.
That bet is SenseNova (日日新), the multimodal system the company has pushed from a lab project into a revenue line.
Key takeaways
- SenseNova (日日新) V6 launched on 10 April 2025 in Shanghai's Xuhui district; the Pro version uses a 620-billion-parameter mixture-of-experts core with native text, image, and video fusion.
- China's Academy of Information and Communications Technology (CAICT) gave SenseNova the top 4+ rating in the first round of its multimodal model assessment, in January 2025.
- SenseTime's 2025 revenue passed RMB 5 billion (≈ US$704 million / HK$5.46 billion), up 33% year on year, a record high.
- Second-half 2025 EBITDA turned positive at RMB 380 million (≈ US$53.5 million / HK$415 million), the first time since the IPO.
- The generative-AI segment was 77% of first-half 2025 revenue, reaching RMB 1.815 billion (≈ US$256 million / HK$1.98 billion), up 73%.
- A companion text-to-image service, MiaoHua (秒画), has served millions of users with more than 100,000 images produced per day.
One model, many modes
SenseNova (日日新) started in 2023 as separate pieces—a chat model, an image model called SenseMirage, and a digital human (数字人) video tool called SenseAvatar. Version 6 tried to merge them. The Pro build uses a 620-billion-parameter mixture-of-experts design that handles text, image, and video as one stream rather than routing between specialist models. A lighter Omni build supports real-time voice and video interaction.
At the launch event, SenseTime showed the model comparing prices from a product photo through speech, and auto-editing a basketball video down to just the made baskets by understanding the footage. Those demos underline the thesis: media value comes from comprehension, not only from synthesis.
The credibility check
Independent signal matters more than launch spectacle. In January 2025, CAICT—a government-linked standards body—ran the first round of its trustworthy multimodal model assessment and gave SenseNova the highest 4+ grade, the first model to reach it. The test covered basic ability, understanding, generation, and specialized tasks across more than thirty items.
SenseTime followed V6 with V6.5 on 30 July 2025, described as moving the system from "tool" to "partner" through tighter real-time text, audio, and video interaction. The Raccoon (小浣熊) productivity family, built on the model, passed 3 million users.
Where generation shows up
Beyond chat, SenseNova (日日新) feeds several media-facing abilities:
- Text-to-image through MiaoHua (秒画), aimed at art and design support, with daily output in the six figures.
- Long video understanding, claimed at ten-minute full-frame parsing, used for auto-clipping and structured summaries.
- Digital human (数字人) video via the earlier SenseAvatar lineage, for spokes-avatar and training content.
- Real-time interaction, where the Omni build takes live voice and video and responds within the same session.
The throughline is "understand then act," which suits enterprises that need analysis (a shelf, a match, a document) more than they need a pretty render.
The 2025 turnaround
The financial story is as relevant as the technical one. SenseTime reported 2025 revenue above RMB 5 billion (≈ US$704 million / HK$5.46 billion), up 33%, its fastest growth in three years, with net loss narrowing 58.6%. Second-half EBITDA reached RMB 380 million (≈ US$53.5 million / HK$415 million), positive for the first time since listing. Generative AI, at 77% of first-half revenue and growing 73%, is now the engine.
That context separates SenseNova (日日新) from pure research showcases: it is tied to a revenue reset, not just a model drop.
Beyond the render
SenseNova's emphasis on comprehension over synthesis is a deliberate positioning. Many buyers in retail, logistics, and education do not need another image generator; they need a system that can read a shelf, a document, or a clip and return a decision. By weighting understanding, SenseTime aims at operational tasks where accuracy and traceability matter more than visual novelty, which is a harder but stickier market than consumer novelty filters.
Honest limitations
SenseNova's media strengths lean toward understanding and editing rather than frontier-quality raw generation; it is not primarily a competitor to dedicated video factories on cinematic output. The CAICT 4+ rating is a Chinese domestic assessment, useful but not the same as an open international leaderboard. Revenue figures are company-reported in the annual results and should be read alongside the still-narrow profit profile—EBITDA turned positive only in the second half. And like all unified models, real-time multimodal interaction depends heavily on latency and device constraints that demos rarely show.
What readers can do now
- If you process large volumes of video or images, trial SenseNova (日日新) V6.5's understanding APIs on a tagging or clipping task before committing to a heavier pipeline.
- For brand visual work, benchmark MiaoHua (秒画) against your current image tool on a fixed prompt set and check consistency across batches.