NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

SenseTime Puts Ten Minutes of Video Into One Model, Then Takes It to Hong Kong

SenseNova V6 pushes long-video reasoning and real-time interaction into a single model, with a subsidy to pull developers in.

2026-11-17 · 867 words · NeuroAI
SenseTime Puts Ten Minutes of Video Into One Model, Then Takes It to Hong Kong

A video model that understands ten minutes of footage sounds like a laboratory curiosity until you watch someone try to use one. Most systems sample a few frames, lose the thread, and quietly miss whatever happened between them. SenseTime's answer, announced at its Tech Day in April 2025, is a large model (大模型) that keeps the whole sequence in view — and then a decision to push it into everyday apps rather than leave it in a paper.

The model family is called SenseNova V6 (日日新 V6). SenseTime describes it as a native multimodal (多模态) mixture-of-experts (混合专家, MoE) foundation model with more than 600 billion parameters. It accepts text, images and video, and it reasons over them rather than merely describing them.

The ten-minute claim

The most concrete capability is what SenseTime calls global memory. The company says V6 is the first model in China to support comprehension and deep reasoning over ten-minute videos at full frame rate. The engineering behind it is less magical than the phrase: visual frames, audio, subtitles and temporal logic are aligned into a unified sequence representation, then compressed with fine-grained cascading compression and content-aware filtering. A ten-minute video, SenseTime says, becomes about 16K tokens while retaining its key semantics.

That compression is the whole trick. Long-context video is normally a memory problem masquerading as a model problem. If you can reduce the input to something the attention mechanism can actually hold, the reasoning becomes possible. SenseTime also says the model can cut highlights automatically, which is a consumer feature built on top of the same representation.

Alongside it sits SenseNova V6 Omni, a lighter full-modal interaction model aimed at real-time use: role-playing, translation, guided tours, picture-book narration and maths explanation. SenseTime says it has been deployed in several industries, including embodied intelligence, and describes it as China's first commercialized full-modality real-time interactive model.

Reasoning with pictures, not just words

The other half of the release is about thinking. SenseTime says V6 was trained on more than 200B high-quality multimodal long chain-of-thought (思维链, CoT) data, with chains up to 64K tokens. Long chains were synthesized and verified through multi-agent collaboration, then used to teach the model to work through image-and-text problems step by step.

The company puts the practical value in document work. In an insurance-claims example, V6 can examine submitted commercial health-insurance materials and flag unnecessary prescriptions or tests, missing documents and incomplete submissions. That is a mundane task, which is precisely the point: a model that can reason over a scanned claim form is closer to a business process than one that wins a maths contest.

Here the brief's fact rule matters. SenseTime's release is full of superlatives — "first in China," "lowest cost in the industry," "most advanced." These are the company's claims, made on its own site, and several are benchmarked against named foreign models. They should be read as vendor assertions. No independent leaderboard result was found to confirm the "first" or "lowest" language.

From Shanghai to a Cantonese app

A notable part of the rollout happened in Hong Kong. SenseTime upgraded SenseChat (商量), its AI assistant for Hong Kong users, on top of V6. The updated app added real-time audio-video interaction with emotional expression, with Mandarin conversation at launch and Cantonese voice calls promised in a later version. It supports traditional Chinese and was made available free on iOS and Android.

That detail is easy to skip, but it is the clearest evidence of strategy. A model advertised on reasoning benchmarks is being shipped through a local-language consumer app. SenseTime's own framing, from chairman and CEO Dr. Xu Li, is blunt about the intended destination: "AI's true purpose is found in our everyday lives."

The voucher

To pull developers toward its platform, SenseTime announced a dedicated subsidy of RMB 100 million (≈US$13.9M / HK$108M), aimed at embodied intelligence and AIGC. The package covers expert consulting, model training and reasoning validation. It is a go-to-market instrument as much as a research announcement: the model is available for trial through the SenseChat web platform, and the subsidy lowers the cost of integrating it.

Honest limitations

  • All capability, ranking and cost claims come from SenseTime's own announcement. The "first in China" and "lowest in the industry" phrases are company assertions, not independently verified results.
  • The ten-minute video figure is a measured capability under SenseTime's compression pipeline, not a guarantee for every video type, resolution or domain.
  • The RMB 100 million voucher is a subsidy with terms likely to change; no detailed eligibility rules were part of the material reviewed here.
  • The app rollout described was at preview stage in the source, and Cantonese voice calling was listed as a future release, not a shipped feature.

What readers can do now

  1. Test long-video claims on your own footage. Frame sampling and full-sequence compression behave very differently on a meeting recording, a lecture and a film trailer.
  2. Treat "first" and "lowest cost" as claims to re-check. Ask which benchmark, which hardware and which price a vendor is comparing before repeating the number.
  3. Watch the local-language app, not just the model card. A model reaching users through a Cantonese assistant tells you more about real deployment than a launch-day benchmark slide.

Related coverage

Cite this article

NeuroAI (脑机网), "SenseTime Puts Ten Minutes of Video Into One Model, Then Takes It to Hong Kong," 17 November 2026, https://neuroai.site/a/na-model-sensenova-v6-video

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories