Bring intelligence to the data.
Bring intelligence to the data. The device is the agent; the model is the product. Through continual learning, models grow with their users.
How it works
The device is the agent; the model is the product. Continual learning is our approach to adapting models to users and their environments.
Learning starts with the data you choose to record or authorize.
Extract a profile from local records and build Neural Imprint learning state on the device.
Base weights remain frozen. Learning state can be saved and restored with a compatible configuration.
Your data physically never leaves your device.
Learning state
Learning state can be saved and restored in later sessions, carrying local learning forward while base weights remain frozen.
Consolidate local records into inference state while base weights remain frozen.
Save state across sessions, check compatibility before activation, and support restore and rollback.
Apple devices use a Metal backend; Intel uses OpenVINO.
In runtime tests, a 9B model completed 200 continuous turns on each of two iPhones.
Evidence and boundariesCross-silicon
We build local inference runtimes using Metal on Apple devices and OpenVINO on Intel Core Ultra.
iPhone / iPad / Mac
Core Ultra · CPU + Arc iGPU
More silicon platforms on the roadmap
We adapt the runtime, operators, and compute graph to Metal and OpenVINO separately.
Intel figures compare runs on the same machine with the same model, an Intel Core Ultra 5 225H. RTF = generation time ÷ audio duration; speech synthesis figures are medians from a 30-minute resident run.
Products
Narrus, Dailyn, and Mealens serve as live validation of our inference runtime. Ururu remains in internal testing.
Research
We are hiring for inference optimization, model interpretability, distributed on-device inference, and full-stack engineering.