Introducing Sequoia 1 Preview

Our first gaze model. It turns the camera a product already has into an eye tracker that learns each person. Available to companies today.

Three giant sequoia trunks, lit orange, in a dark forest.

Photo: Miguel Vieira, CC BY 4.0, via iNaturalist

Today we're releasing Sequoia 1 Preview, the first model in our Sequoia family. It gives any product with a camera accurate eye tracking, without dedicated hardware, and it gets more accurate the longer each person uses it.

Learns each person

Faces, cameras and seats all differ. Instead of one model for everyone, Sequoia fits a model to each person on their own device, then keeps sharpening it with every action they confirm: a click, a tap, a press of a button. After twenty of them, gaze error falls by 27%.

Seated, first use
1.65°
Seated, after 20 uses
1.20°
Moving, after 20 uses
1.78°
Precision
0.72°

How it compares

On first use, Sequoia 1 Preview is level with Tobii Nexus, Tobii's webcam eye tracker. After twenty uses it is 27% more accurate, and its precision is within 0.04° of Tobii's dedicated infrared tracker.

Accuracy on a webcamGaze error in degrees. Lower is better.
View data
Sequoia 1 Preview, after 20 uses1.20°
Sequoia 1 Preview, first use1.65°
Tobii Nexus1.64°
WebGazer4.17°
PrecisionSpread of repeated estimates, in degrees. Lower is better.
View data
Sequoia 1 Preview, webcam0.72°
Tobii Nexus, webcam0.85°
Tobii Pro Nano, infrared hardware0.68°

Tobii Nexus and Tobii Pro Nano: laboratory results with 101 participants from Toward Scalable Eye-Tracking: Validating the Tobii Nexus Webcam-Based System (2026). WebGazer: Papoutsaki et al., IJCAI 2016. Sequoia: llocus measurements across 15 hardware setups in ordinary rooms and changing light.

Built for the car first

Since July 2026, every new car, van, truck and bus sold in the EU must warn a driver who looks away from the road, so every new vehicle now carries a camera on its driver. A car has a few drivers for years, and they show where they are looking all the time: every touch of the screen, every press of a wheel button. That is the setting Sequoia was designed for.

The same holds in cockpits and simulators, on accessibility devices, in research and on industrial controls.

Small enough to share the machine

Sequoia needs no GPU and no dedicated hardware. We optimized it with Kestrel, our engine for running neural networks on the chips products already have. Camera frames, faces and each person's model stay on the device.

Availability

Sequoia 1 Preview is available to companies today. If you're building a car, a cockpit or a device that should know where people look, talk to us.

More from llocus