Z-Next system card

System overview

Two systems use Z-Next. Both start from the same model: the product variant of Z-Next 1B, an English language model with no attention layer and no KV cache, whose state per conversation is fixed at about 5.2 MiB. ZChat serves the product variant as it is. Z-Town starts from it and tunes it on the village's own record after each village night; the benchmark figures on this card are for the product variant itself.

The demo is switched on by hand and has been offline before, so it may be off when you visit: open the live demo. The measured results are in the evaluation section of the model card.

In Z-Town, the world's rules decide which moves are allowed, the model chooses among them, and each resident keeps its own fixed-size state. No resident follows a script. More on the Z-Town case page.

Model and versions

The demo engine starts the product variant by default. At launch the demo could be started with either of two chat variants, and some of our run records from 2 Sep and 10 Oct 2026 list the later chat variant as the one served. We have not yet reconciled those records. The four variants and their results are on the model card.

Capabilities and measured evaluations

Z-Town residents learn from what they live: after each village night the model is tuned on the village's own record. We are still measuring how much that changes their behaviour. The model card has the full evaluation, including the comparison with the transformer twin and its recipe caveat.

Known limitations

• Chat quality is not product grade. The product variant failed its acceptance gates (see the table above).
• It makes things up, often, as expected at 1 billion parameters and 10 billion training tokens. Do not trust any fact it states.
• English only. Other languages were not trained for and are not evaluated.
• Long documents. A fixed-size state does not mean it recalls everything. Recall over long inputs is near chance. In one earlier test through the demo a planted fact was lost beyond a few thousand tokens. In informal demo tests on 11 Nov 2026 a short code came back from long notes in most runs, and other facts often did not (see the model card).
• Reasoning. Multi-step questions essentially fail.
• Speed. Roughly 48 tokens per second single-stream on one RTX 5070 Ti, unoptimised. Each step is slower than in a comparable transformer, and batched serving across many live sessions has not been measured.
• Availability. The demo runs on one GPU, is switched on by hand and has been offline for days at a time.
• Z-Town. The village is open to visitors aged 18 and over. Visitors can talk with residents; replies come from a 1B model and can be wrong.

Risks, misuse and mitigations

The demo is for looking at how the model behaves. It is not for advice of any kind, and not for any decision about people.

Anyone using the demo can prompt the model into producing text that is false or harmful. Nothing in the model prevents it.

Safety work done and not done

Done

• Pre-training data passed through adult-content and spam filters.
• Acceptance gates for chat and tool use written before training and not moved afterwards. The product variant's result, 14 of 17, is reported as a fail.
• A checksum check on the weights before the demo engine starts.
• CAPTCHA and proof-of-work at demo entry.
• Public documentation of limitations and negative results, on this card and on the model card.

Not done

• No RLHF or other preference tuning.
• No red-teaming.
• No content filter on input or output.
• No bias or toxicity evaluation.
• No evaluation in any language other than English.
• No independent review or reproduction.

Data handling in the demo

What we can state: the model runs on a GPU we operate ourselves, the public web server holds no model, and entry is gated by a CAPTCHA and a proof-of-work check. For what the demo collects, what is logged and how long anything is kept, see the privacy notice.

Do not type personal, confidential or sensitive information into the demo. The same applies to anything you type to a resident in Z-Town.

Change log

Contact for issues

To report a harmful output, a security problem, or a number on this card that looks wrong, write to [email protected]. Tell us the date and time and what happened; please leave out personal details. Corrections go into the change log above with the date and what changed.