Nexith Core is now serving on our own sovereign GPU cluster — cache hits save up to 90%, OpenAI-compatible.Learn more →

Engineering notes

How Nexith is built.

Short, honest write-ups on the systems behind the model — caching, memory, routing, and privacy. No benchmarks we can’t reproduce; just the engineering.

Systems

The semantic cache that saves up to 90% on a hit

How we use local embeddings and cosine recall to serve repeated requests from cache — turning an upstream call into a low-cost, near-instant response.

Read the note

Systems

Giving the model a memory that survives the session

Pinned facts and auto-extracted context, scoped per user and app, retrieved at request time. What it takes to make “remembering” cheap and predictable.

Read the note

Inference

Routing a request to the right place, cheaply

A lightweight classifier decides how each request should be handled before it ever reaches the model. The design trade-offs behind keeping that step fast.

Read the note

Privacy

One egress gate for everything that leaves

Every response — streaming, non-streaming, or served from cache — passes a single allowlist gate on the way out. Why a single chokepoint beats scattered checks.

Read the note

Want the full detail?

The research page goes deeper on the systems that make Nexith fast and cheap to build on.

Read the research →