The Edge
The whole thing runs on a
$250 computer.
Not a cluster. A single board on a desk, running an agent that remembers ten million tokens, wakes itself up, and works while nobody is watching. The interesting engineering is what fits, not what scales.
Measured on the device
Recall across a 10-million-token context
two seeds, 92/92/95 sweep at seed 42 — never quote the 95 alone
88–95%
Median retrieval latency
on the Jetson itself, not a cloud proxy
756 ms
Distilled specialist model, quantised
Q8_0 GGUF, certified on-device run
12.3 MB
Caveat — these are our own benchmarks on our own hardware, published with their methods and their seeds. Recall is a range because two seeds disagreed, and the honest number is the range. The benchmark write-ups →
Why a small machine is the point
An agent that lives on hardware you own has properties a hosted one cannot have. It keeps working when the internet doesn't. Its memory is a file on your disk rather than a row in someone else's database. And its running cost is electricity.
The constraint is also the research programme. Making a capable agent fit in 8 GB forces the questions that matter — what does it actually need to remember, which decisions need a large model and which need a small one that answers in milliseconds.
The distillery
Most of what an agent does every hour is reflex, not reasoning: classify this, route that, decide whether this is worth waking up for. Those calls do not need a hundred-billion-parameter model, and paying one to make them is the largest avoidable cost in the whole system.
So we distil specialists — small models trained on one bounded job, certified on the device they run on. The benchmark artifacts are their first publications, and the method is the product: we show the runs, the seeds and the failures.
Caveat — distillation targets reflexes, not judgement. Anything requiring actual reasoning still goes to a frontier model, and we say which did what.
Read the work
The field notes are the research wing — written up in full, including the runs that disagreed with each other.
Field Notes →The product side
The same intelligence, pointed at a business instead of a bench.
The workforce →