Assistant updates through your OTA
Ship new assistant versions over your own update channel or ours, signed, staged and reversible.
Device makers (OEM) · Build, tune and ship
One model, sized for every SKU
The same assistant is compressed for each device’s chip and memory, with an accuracy report per SKU.
- Desktop
- Works offline
Desktop assistantOn this device · no internet needed
How much accuracy do we keep on the entry-level model?
The accuracy report for that SKU shows the result on your own test questions, next to the flagship. Example figures only until your pilot runs.
Can the entry model run the full assistant?
The SKU report shows which features fit its memory. Larger features can stay on higher models.
How long does a new SKU take?
It depends on the chip. Adding a SKU means compressing and testing, which is planned in the tuning phase.
The problem
A range spans flagship and entry-level devices with very different chips and memory. Building a separate assistant for each is slow and expensive. Teams need to know what quality each model will deliver before committing.
Who uses it
Device maker product and engineering teams planning a range of models.
What you train it on
- Your assistant content
- Device specifications per SKU
- Your test question set
How it works on the device
- 1
One trained model
The assistant is trained once on your content.
- 2
Compressed per SKU
It is fitted to each SKU’s NPU, GPU or CPU and memory budget.
- 3
Accuracy report per SKU
Each build is tested on your own test questions, so you see what each device keeps.
Why it runs on the device
One assistant for the range
Same behaviour across flagship and entry models.
Decisions on evidence
Accuracy per SKU is measured on your own questions.
Faster launches
New SKUs are added by compressing, not rebuilding.
Questions
Common phone, PC and automotive NPUs, GPUs and CPUs. Your specific chips are checked in evaluation.
Yes. Every SKU build comes with its accuracy report on your test questions.
Yes, and we recommend it. Results on your own questions matter most.
Related use cases
All use casesPut AI where your work happens.
Pick one team and one use case. We’ll have it running on your own devices in weeks.