Xailient, which makes edge AI face recognition software, said on Nov. 17, 2026 that its newest model came first for accuracy in the Border-Kiosk test, one strand of the Face Recognition Technology Evaluation (FRTE) programme at NIST (National Institute of Standards and Technology). The ranking covers a group of 106 developers whose algorithms need 500 milliseconds or less to create a template. The claim comes from the company's own announcement, distributed through PR Newswire.

Advertisement

Why the company says this test matters

Xailient describes Border-Kiosk as the only NIST 1:1 evaluation built on still images of people who are not cooperating with the camera. Subjects tap a touchscreen mounted below the lens, so their faces are often tilted downward, sometimes past 45 degrees. Those shots are then compared with photos taken at border crossings, so neither picture in a pair is posed.

According to Xailient, on-device systems in the NIST setup are capped at 1 GB of RAM and must respond within 500ms, whereas server or hub systems have no fixed ceiling. NIST evaluated xailient_001, the same model Xailient says it ships on edge hardware offering about 1 TOPS of NPU compute.

Reported results

The company says that with the false match rate fixed at 1 in 100,000, its model got 95.4% of the genuine pairs right in this test, the top score in the category. At a stricter setting of 1 in 1,000,000, it reports a result equal to 94.0%. For the easier Border-Border test, Xailient cites a 0.55% false non-match rate at 1 in 1,000,000, which it puts at approximately 99.45% genuine-match accuracy.

Xailient says NIST measured its template generation at 497 milliseconds. In a separate comparison against six major rivals, carried out by the company rather than NIST, it says the model ran 1.6× faster than close competitors and missed 12% fewer matches. Its memory footprint was 901 MB and its configuration data 280 MB, which Xailient says were the lowest in that group.

Cost, privacy and live testing

Xailient argues that lighter hardware demands cut device costs, citing RAM prices up 400% since 2024 and Flash memory up 250%. Chief executive Lars Oleson said the results show real-world accuracy within a far smaller compute envelope. Mallika Patel, VP of Product, said that matching on the camera means bystander data is not gathered.

Advertisement

The company also pointed to recent ISO live-subject testing in which 250 diverse, moving people were checked in real time against 10,000 registered candidates, producing 99.5% successful recognition and no false matches. It notes that its production system draws on multiple frames, unlike the single-pair NIST benchmark.

NIST's published results are not an endorsement by the agency of any company or product. The agency says FRTE runs on a continuous basis, and under its rules each participant can submit no more than once every four calendar months.

Sources