AI & ML interests

None defined yet.

Recent Activity

olumideola  updated a collection about 17 hours ago
Diacnet 2
olumideola  updated a Space about 17 hours ago
olaverse/diacnet-1.0
olumideola  updated a Space about 17 hours ago
olaverse/LID_Models
View all activity

Organization Card

olaverse

Olaverse Lab

Small, task-specific open models. Open-source NLP infrastructure for underrepresented languages, built African-first.

We build one focused model at a time, each doing one job well and cheap to run in production. Our flagship is PurpleMIST, a family of decision models that go head to head with models hundreds of times their size. Alongside it are specialists that identify the language, restore its marks, turn documents into structured data and generate questions from it. Every model has open weights.

The language-ID and diacritic models run locally through our library, with no API keys, no sign-ups and no rate limits:

pip install olaverse

Website · Docs · GitHub · PyPI · X · LinkedIn · YouTube · Reddit

Latest releases

Family Models What it does
PurpleMIST (flagship) New: Mini-2.0 (4.5B) · Flash-1.0 (7.9B) · Mini-1.0 (1.9B) Calibrated decisions (choice, yes/no, score) about a message, record or image. Mini-2.0 matches Flash-1.0 on the Decision Index at half the size. It also reads images, works in 19 languages and takes inputs over 100k tokens without truncating them. Flash matches DeepSeek-V4-Pro's accuracy in English and across eight African languages, at about 1/200 of its size
MIST-Extract 1.0 (4B) · Flash-1.0 (2B) Turns documents (text, photos or scans) into JSON, with a calibrated confidence for every field. The strongest model we tested on African-language documents, and 91.8% of held-out English documents extracted fully right
diacnet / diactag diacnet-2.0 · diacnet-mini-2.0 · diactag-2.0 Restores tone marks, diacritics and Arabic vowel marks in 11 languages
LID-608 lid-neural-608 · lid-lite-608 (37 MB) Identifies 608 languages, built African-first

Benchmarks

Benchmark What it measures
African Typed Decisions System One decisions in eight African languages, with a leaderboard
DiacBench Diacritic restoration
Decision Index results Our complete runs on the Decision Index public text and vision suites

Release notes land on Hugging Face first: follow olaverse to catch every release.

Work with us

Building for these languages, or want to use these models in production? Get in touch through olaverse.co.uk or open a discussion on any model.

Olaverse Lab: building AI that speaks your language.