# Antilattice

> Antilattice builds Latent Graph Attention — the first sub-quadratic attention
> architecture. It replaces dense attention's O(n²) pair grid with a learned
> token graph, routing compute and memory only along the edges that matter.

## Key facts

- **Architecture**: Latent Graph Attention — sub-quadratic, near-linear cost scaling (O(n²) → O(n) vs. dense attention)
- **Quality**: dense-matched, token-for-token — no loss in reasoning quality vs. dense quadratic attention
- **Speed**: up to 10× faster inference at 512K context, same hardware
- **Context**: first-ever 5M-token context window, served at production cost
- **On-device model (Antilattice Edge)**: 1B parameters at full fp16 precision, 256K-token context in 4GB of RAM, fully offline, 0 bytes leave the device
- **Company**: Antilattice, Hyderabad, India — hello@antilattice.com

## Why it matters

Dense attention computes every token pair: compute quadruples each time context
doubles, and the KV cache exhausts GPU memory long before compute does. Antilattice
learns which token relationships carry meaning and keeps only those — the KV cache
collapses from 51.5GB to 0.5GB at 256K tokens for a 1B model, which is what lets a
full-precision model fit in a phone's 4GB memory budget.

## Pages

- [Antilattice — main page](https://www.antilattice.com/): the architecture, benchmarks (10× throughput at 512K, still serving at 5M), use cases, and the bottleneck of dense attention
- [Antilattice Edge — on-device](https://www.antilattice.com/on-device): the 1B fp16 on-device model — the memory math (256K in 4GB), on-device benchmarks, and the beta