This demo runs on one workstation GPU. The check below keeps automated
traffic from taking the seats.
No cookies, no third-party scripts. What you type is processed to produce the
reply and is not stored; the session is discarded when you leave.
We count visits in aggregate only — page, country and referring site,
with no cookies and no IP address stored.
Tetracta Z1 — the conversation grows, the memory doesn't.
Z1 is a 1-billion-parameter language model trained on 10 billion tokens with
no attention and no KV cache. Its session memory is a fixed
5.227 MiB — the same at turn 1 and at turn 100.
Beside it runs a standard transformer trained on
identical data, whose memory grows with every token. Ask them both the same
thing. One workstation GPU serves
hundreds of these conversations at once, because each one costs
the same fixed amount no matter how long it runs.
1B parameters · 10B tokensconstant-memory architecturesingle GPU, split in halfresearch preview — it makes mistakes
Concurrent seats · Z1
–
on its half of the GPU
Concurrent seats · transformer
–
same GPU, same half
Ratio
–
more people, same silicon
Right now
connecting…
Tetracta Z1
constant-memory architecture · session state fixed
session state 0.000 MiB0/0 seats used
a standard transformer
trained on identical data · KV cache grows with context
KV reserved 0 MiB0/0 seats used
Sampling — applies to both models equally, so the comparison stays fair.
Your settings are yours alone; they do not affect anyone else's session.
Using the default settings.
Paste as much as you like — up to 32,000 tokens in a
single message. No message limit. Sessions idle for 10 minutes are released.
Resetting frees your own seat only.
Honest note. Z1 is a research preview at 1B parameters. It will get facts
wrong, mangle arithmetic and occasionally answer confidently about things it does
not know. On retrieval it is honest-but-limited: it will pull a specific detail back
out of a page or two of text, but past a few thousand tokens that retrieval falls
away — the architecture carries the context for free, the 1B model has not yet
been taught to search it. That gap is the next training run, not a limit of the design. What this demo is showing is not encyclopaedic accuracy — it is what the
memory does while you keep talking. Full evaluation appears in the upcoming paper.