Technology03
Edge inference
reasoning that lives on the device
Cloud round-trips are a luxury a moving robot doesn't have. XINK CORE runs LLM and VLM reasoning where the data is captured — private, low-latency, and still working when the network isn't.
- 4.6 TOPS
- NPU on the eCV SoC
- 1B
- params — LLM running in live demos
- 28 nm
- next-gen LLM IC in tape-out
01
Why the edge wins
Latency: decisions in milliseconds, not round-trips. Privacy: video and voice never leave the device. Resilience: inference keeps running offline. For machines that act in the physical world, edge inference isn't an optimization — it's a requirement.
02
Edge-cloud hybrid, by design
Small models reason on-device — a quantized Gemma 3 1B handles intent parsing, privacy filtering, and instant interaction — while the cloud orchestrates heavy lifting: planning, external APIs, large-model support. Secure transport, two-way collaboration, each side doing what it's best at.
03
Language meets vision
Give an LLM agent a camera — captioning, visual question answering, phrase grounding, detection — and a robot can learn its surroundings and adapt to objects it has never seen. That's what turns automation into autonomy: environments no longer need to be predictable.
04
Next: a processor built for language
The upcoming 28 nm LLM IC targets low-context inference with efficient prefill and decode — real-time intelligence and context awareness at edge power. It completes the Sense-and-React architecture: perception and cognition as a single edge reflex.
Start with a dev kit, scale to a fleet.
Get XINK SENSE on your bench this week, or talk to us about the full stack.