What Seph can and cannot do

Seph is a small language model. The default is 108 million parameters; the larger option is 596 million. For comparison, the assistants most people have used are somewhere between a hundred and a thousand times bigger. That difference is not a detail — it decides what is possible.

This page is specific on purpose. "May produce inaccurate information" is true of everything and tells you nothing about when to trust a given answer.

Do not trust it for

• Arithmetic it does in its head. Multi-digit multiplication is unreliable at this size. Seph has a calculator tool and is much better when it uses it — if you see a calculator chip above the answer, the number came from real arithmetic; if you do not, it did not.
• Quotations reproduced from memory. Verbatim recall drifts: a quoted verse or passage can come back with words changed. When Seph looks a verse up with a tool, the text is exact, because it comes from a database rather than from the model. Again: look for the chip.
• Anything where being wrong is costly. Medical, legal, financial, or safety decisions. Not because of a liability formula, but because a model this size confabulates confidently and you should not be the test case.
• Recent events, unless the answer shows a search chip. The weights have a training cutoff and no awareness of it.

What it is genuinely reasonable at

• Deciding which tool to call for a question, and with what arguments — a routing task that suits a small model well.
• Ordinary conversation, rephrasing, and summarising text you provide.
• Questions answerable from a source it has just retrieved and can quote.

A limitation we have measured and not yet solved

There is a real gap between Seph retrieving the right information and Seph using it. It will often run the correct search, get a page that contains the answer, and then write a reply that does not draw the answer out of it — particularly when the question needs two retrieved facts compared against each other. We can measure this reliably; it has not gone away with additional training at this size, and the evidence points to model capacity rather than something a prompt fixes.

Practically: when an answer matters, read the sources in the chips rather than only the summary. The retrieved page is often right when the summary of it is not.

Context length

Seph reads about 2,048 tokens at a time — roughly 1,500 words including the conversation so far and any retrieved text. Past that, the oldest turns are dropped. When that happens the app tells you exactly how many turns were left out rather than silently truncating.

Speed

Replies are generated on a CPU, not a GPU. A reply takes seconds rather than being instant, and a search adds more. That is the cost of running a model that does not need a datacentre.

Why publish this

Seph is a research project as much as a product, and the interesting result is not "a small model can do everything" — it is finding out precisely where the ceiling sits. Telling you where it fails is the honest version of that, and it is more useful than a disclaimer.

If Seph gets something wrong, the thumbs-down button under any reply is the most useful thing you can do with it — those reports are read.