Blog Product updates

Orthos 1.1: Knolt for every camera in the branch

Orthos 1.1 is our light model for the everyday cameras: faster than Talos 1.1, with fewer false alarms.

Knolt engineering 2 min read

A bank branch has a few cameras where the worst can happen: the teller line, the ATMs, the vault, the front door. It also has many more that watch ordinary space: the parking lot, the corridors, the lobby, the back door. Until now, Knolt watched all of them with the same model. Orthos 1.1 changes that.

Two models, one generation

Talos 1.1 stays our flagship. It catches the most incidents, and it watches the cameras where a robbery or an ATM attack would happen.

Orthos 1.1 is new. It is a lighter model for the everyday cameras, built to cover many cameras on less hardware. The number is the generation: Talos 1.1 and Orthos 1.1 ship together.

What Orthos 1.1 is good at

On Security-Bench 1.3, our test of one question (is this a security incident?):

  • Fast. It answers a query in 0.09 s, against 0.15 s for Talos 1.1.
  • Few false alarms. 97% of its alerts were real, and it left 98% of ordinary moments alone. That matters on busy cameras, where every false alert costs your team time.
  • Ranks well. Its ranking quality (AUC) is 92%, the highest of the models we tested.

Talos 1.1 and Orthos 1.1 on Security-Bench 1.3. Incidents caught: 75% and 59%. Alerts that were real: 88% and 97%. Ordinary moments left alone: 91% and 98%. Ranking quality (AUC): 91% and 92%. Time per query: 0.15 s and 0.09 s.

Where it is not the right model

Orthos 1.1 catches fewer incidents than Talos 1.1: 59% on Security-Bench 1.3, against 75%. It is weakest on robbery and theft. That is why it does not watch the teller line, the ATMs or the vault. Those cameras stay with Talos 1.1.

What stays the same

Orthos 1.1 runs where Talos does: on your own GPU, in your branch or your own cloud. Your footage stays yours. A flag is a lead for your team to review, never a verdict. A person decides.

The full comparison

Every model saw the same frames and got the same question.

All five models on Security-Bench 1.3. Correct calls, incidents caught, alerts that were real, ranking quality and time per query. Talos 1.1: 83%, 75%, 88%, 91%, 0.15 s. Orthos 1.1: 80%, 59%, 97%, 92%, 0.09 s. Claude Opus 5.5: 80%, 69%, 85%, 87%, 1.71 s. GPT-6.1 Sol: 74%, 54%, 86%, 85%, 2.81 s. Claude Sonnet 5.5: 72%, 51%, 86%, 84%, 1.91 s.