Introduction
We’re introducing Tessera, a very small chat model built for research. It has 49,995 parameters and typically writes a reply in about 1 to 5 milliseconds.
Tessera exists to answer one question: how much conversation can a model this small actually hold? It’s an experiment we’re sharing, not an assistant to rely on. A tessera is one small tile in a mosaic. No single tile shows you much, but enough of them in the right places make a picture.
Here’s where it stands:
Practice tests. Tessera passes 126 of our 164 practice messages, up from 111 for the older version. It keeps everything the older version got right and improves on every social test.
Why this version. A newer version scores higher, at 138. But on a wider check, most of the replies it changed turned into a flat “okay”. Tessera gives fuller answers, so it’s the one we’re releasing.
Speed. Tessera writes a typical reply in about 1 to 5 milliseconds, long before you could blink.
Research only. Tessera can be confidently wrong, and it isn’t meant for everyday use. For anything that matters, use the default model in Sennoric Chat.
Performance
Each practice test is a short set of messages with a pass rule, such as “answer a greeting with a greeting”. We compare Tessera with the older version it replaces, and with a newer version that scored highest in the same round of experiments.
| Testmessages in the test | Tesserathis release | Older versionthe one before Tessera | Newer versionhigher score, flatter replies |
|---|---|---|---|
| Greetings16 messages | 13 | 11 | 16 |
| Thanks and appreciation16 messages | 15 | 13 | 16 |
| Goodbyes16 messages | 9 | 6 | 13 |
| Replies that need context16 messages | 9 | 6 | 12 |
| Wellbeing questions16 messages | 12 | 10 | 12 |
| Acknowledgments like “okay”16 messages | 8 | 7 | 9 |
| Reworded facts20 messages | 12 | 10 | 12 |
| Arithmetic, copying, names48 messages 1 | 48 | 48 | 48 |
| Total164 messages | 126 | 111 | 138 |
Scores are messages passed. The shaded cell in each row is the highest score; Tessera’s column is highlighted throughout. These tests were used during development, so they show progress rather than general ability.
1Arithmetic, copying, names: three separate tests of 16 messages each, combined here. Every version passes all 48. ↩
Set out message by message, the gains are easy to see. Switch between the versions to watch where each one passes.
Why not the higher score?
The newer version wins on the practice tests, so we ran it against 8,034 extra messages and compared its replies with the older version’s. It changed 837 of them, and 711 of those changes were to just “okay”. A thank-you or a “how’s your day” deserves more than that.
of the newer version’s changed replies were just “okay”. The squares are the 837 replies it changed; the light ones are the “okay”s.
Tessera hasn’t been through the same check yet, so we can’t say how often it changes the older version’s replies. What we can say is that it keeps every practice message the older version passed, and it doesn’t fall back on bland answers in the tests we have. We’ll publish that result once the check has run.
A research model
Tessera is small on purpose. Here is every one of its parameters, one square each.
That size is the point of the research, and it’s also the limit. Tessera handles greetings, thanks, small talk, simple sums and remembering your name. Beyond that it guesses, and its guesses can sound very sure of themselves.
Not for everyday use
Please don’t rely on Tessera for facts or advice. It wasn’t built to know things, and it loses track of a conversation after a couple of exchanges. Treat it as something to explore and learn from, the way you’d try out a prototype.
If you need an answer you can trust, use the default model in Sennoric Chat. Tessera is there for curiosity, not for work.
Try it
Say hello to Tessera below. Short messages suit it best, and a greeting or a simple sum is a good start.
Pick a message below or type your own.
When it gets something wrong, it usually gets it wrong confidently. That comes with being this small, and it’s why Tessera is a research release.
Notes
How the tests were run
- Practice tests: eight sets of short messages, 164 in all, each with a pass rule. They were used during development, so the scores show progress, not performance on new messages.
- Arithmetic, copying, names: three separate tests of 16 messages each, reported together. Every version here passes all 48.
- The “okay” check: 8,034 extra test messages. The newer version changed 837 replies compared with the older one, and 711 of those became just “okay”. The same check hasn’t been run on Tessera yet.
- Scope: none of these results shows general ability. Tessera is a small research model tested on short messages.