On-device intelligence
The fastest AI ever built for the device in your pocket.
A 1B fp16 model holding 256K-token context in 4GB of RAM. It runs entirely on-device — no network, no queue, no bill.
262,144 tokens resident · 3.1GB peak footprint · full fp16 weights
Summarize every contract I've signed this year.
256K
Token context window, resident in RAM
4GB
Total memory budget — weights, cache, runtime
1B fp16
Full-precision weights. No int4 shortcuts
0 bytes
Sent to the cloud. Ever
The memory math
256K context. 4GB of RAM. It fits.
A 1B fp16 model is 2GB of weights. Dense attention adds ~192KB of KV per token — 51.5GB at full context, thirteen times the entire budget. Latent Graph Attention routes the cache along learned edges at ~2KB per token, and the whole footprint lands at 3.1GB.
The cache alone outweighs the phone thirteen times over. Squeeze the same dense model into one budget and you fit roughly 8K tokens — Edge holds 256K, a 32× larger window.
How it works
Attention that fits.
Dense
Every pair. Every byte.
Standard attention stores KV for every token it sees. At 256K, a 1B model's cache alone would pass 51GB — no phone was ever going to hold that.
192KB / token
Learned
The model finds what matters.
Latent Graph Attention builds a token graph as it reads — the relationships that carry meaning, and nothing else.
graph built on the fly
Routed
The cache collapses to its edges.
KV survives only along routed edges. The cache shrinks from 51.5GB to 0.5GB, and 256K context fits beside full fp16 weights.
~2KB / token
Measured on-device
78% of short-context speed — at 256K.
Internal benchmarking on a 2026 flagship smartphone with a 4GB memory budget, normalized to short-context generation speed. Dense fp16 KV exhausts the budget before 16K tokens; Antilattice Edge keeps serving to 256K.
Result
32×
more context than dense fp16 fits in the same 4GB budget — and still generating at 78% of short-context speed, with the radio off.
78 tok/sWhy on-device
The cloud was never the only option.
Private by architecture
Prompts, files, and history never leave the hardware. Privacy is a property of the system, not a policy.
Offline is a feature
Airplane mode, dead zones, remote field sites — full-context reasoning with the radio off.
Free at the margin
No per-token bills, no rate limits, no keys to rotate. Once it's on device, inference costs battery — nothing else.
Full precision, not int4
Quantization trades quality for space. Routed KV makes that trade unnecessary — fp16 on a phone.
Get started
Same intelligence. None of the datacenter.
We build frontier models that live entirely on the device. If your product needs intelligence that never leaves the hardware, we want to hear from you.