Hikvision's Guanlan boosts video AI accuracy

1/ Hikvision said on Sept. 8 that it had upgraded applications built on its Guanlan large-scale AI models, pitching video systems that can interpret scenes, search past footage and respond to natural-language prompts. (prnewswire.com) 2/ The company’s framing is straightforward: better accuracy, lower operating cost and less manual review in security and monitoring setups. Hikvision said the newer Guanlan-based systems can “understand” what they are seeing, what has already been recorded and what a user is asking for. (prnewswire.com) 3/ The most concrete claim in the release is at the perimeter. Hikvision said DeepinViewX cameras running the model cut false alarms by 90%, detect at twice the range of traditional cameras and halve repeated alarms on the same target. Those are company claims, not independently verified benchmarks. (prnewswire.com) 4/ Hikvision also said scenario-specific analytics typically reduce false alarms by 70% to 90% in specialized settings. The examples it gave were PPE detection on construction sites, fall detection in care homes, smoking detection at gas stations, forklift detection in logistics parks, and early fire and smoke detection in mining. (prnewswire.com) 5/ What matters here is not just “AI in cameras.” Hikvision is describing a system that blends edge devices, stored video and natural-language interaction so an operator can move from passive monitoring toward query-based review. In plain terms: less scrubbing through footage, more asking the system what happened. (prnewswire.com) 6/ That fits the architecture Hikvision laid out when it introduced Guanlan in 2025. At launch, the company described a three-tier stack: foundation models for vision, language and multimodal tasks; industry models in the middle; and task models on top for specific applications. (prnewswire.com) 7/ The Sept. 8 release suggests Hikvision is now trying to show that architecture in deployed use cases rather than only as a platform story. The emphasis is on concrete jobs where false positives are expensive or where operators need fast retrieval from large video archives. (prnewswire.com) 8/ There is also a cost angle beyond detection. In May, Hikvision launched Guanlan Encoding, an AI-powered compression tool it said could cut storage costs by 30% to 50% on average while remaining compatible with H.265 decoders and third-party devices. That matters because video AI economics are shaped as much by storage and retention as by model accuracy. (prnewswire.com) 9/ Read together, the two announcements show Hikvision pushing Guanlan as both an accuracy layer and an efficiency layer: fewer nuisance alerts on the front end, less storage burden on the back end. That is the vendor’s pitch to buyers running large surveillance and monitoring estates. (prnewswire.com) 10/ The bigger takeaway is narrower than the marketing language. Hikvision is arguing that video systems are becoming more conversational and more retrieval-oriented, not just better at spotting objects frame by frame. If that claim holds in live deployments, the operational change is in how staff search, verify and act on recorded events. (prnewswire.com) 11/ One caution: nearly all of the specific performance figures in circulation here come from Hikvision’s own materials. The release does not, in the excerpts available, cite third-party testing, customer names or standardized external benchmarks for those gains. (prnewswire.com) 12/ So the story on Sept. 8 is less “new model launch” than “commercial proof-point push.” More than a year after introducing Guanlan, Hikvision is using application-level claims in perimeter security, safety compliance and industrial monitoring to show where it says the models are already paying off. (prnewswire.com)