Generate a short, human title from a chat message or document in ~300M parameters — small enough to run on the same box as your app.
Every chat product ends up with a sidebar full of "New chat, New chat, New chat". Every document tool ends up with "Untitled (3)". Users will not name things, so the list becomes unusable at exactly the point it becomes valuable.
The usual fix is calling a large LLM for a five-word title — a per-token bill and a network round trip for a piece of throwaway UI text, on every single new thread.
No prompt template, no system message, no language tag. mist-tg-0.3b takes the raw text and emits a title, so there is no prompt to drift and nothing to maintain.
mist-tg-0.3bFine-tuned from google/byt5-small on 115K real chat messages and their titles. Byte-level seq2seq, 256 bytes in, 64 bytes out — it fits alongside your application server instead of behind an API bill.
mist-tg-0.3bTitles come back short enough to use unedited in a sidebar. Batch a backlog of untitled threads in one forward pass, and regenerate on the second or third message when the first turns out to be a greeting.
The Olaverse SDK wraps the models with sane defaults. If you would rather not add a dependency, the second tab is the same pipeline in plain transformers.
from olaverse import MISTTitleGenerator
titler = MISTTitleGenerator() # mist-tg-0.3b, ~300M params
titler.load()
titler.generate("My laptop keeps freezing every time I open more than five browser tabs, any idea why?")
# → 'Laptop Freezing Impact'
# Clear a backlog of untitled threads in one forward pass
titler.generate_batch(["How do I center a div in CSS?", "What makes Yoruba tonal?"])Every figure on this page comes from the published model cards, and every limitation is one the cards state themselves. Download the weights and check us.
Generate natural, search-style questions from any passage across 25 languages — for quizzes, reading comprehension, help-centre FAQs, and assessment banks.
/ Routing & moderationDetect the language of support tickets, chat messages, queries, and documents across 25 languages — including the African languages most detectors get wrong.
The weights are open, so you can build it yourself this afternoon. If you would rather we tuned it to your domain, evaluated it properly, and handed it over, tell us what you are working on.