Olaverse Lab
Small, task-specific open models — with a core focus on African & Nigerian languages.
We believe the future of open AI isn't only bigger models — it's specialist models: compact, efficient, and sharply focused on one job. We build the full stack — language tools, retrieval, generation, and vision — and release it openly, with a mission to make African languages first-class citizens of AI.
30+ open models · Yorùbá · Igbo · Hausa · Nigerian Pidgin — and beyond
🔥 Featured
| Model | Task | Why it matters |
|---|---|---|
| diacnet-1.0 | ✍️ Diacritic restoration | Our most-used model — restores diacritics/tone marks across Yorùbá, Igbo, Hausa & more. A ByT5 model doing a job almost nobody else does for these languages |
| naija-embed-base | 🧭 Nigerian embeddings | Cross-lingual sentence embeddings for Hausa, Yorùbá & Igbo — the foundation for local-language search & RAG |
| mist-qg-1.5b | ❓ Question generation | Generates questions from any passage in 25+ languages · demo Space |
| lid-neural-5.1 | 🌍 Language ID | Identifies Nigerian languages — built on our own ModernBERT encoder · live demo |
🌍 Nigerian & African Language Stack
Over 2,000 of the world's languages are African, yet they remain nearly invisible in AI. We're changing that — starting with Nigeria, one focused model at a time.
✍️ DiacNet — diacritic & tone restoration
Restoring the marks that make Nigerian-language text readable, searchable, and machine-usable.
- diacnet-1.0 — flagship multilingual ByT5 restorer (Yorùbá, Igbo, Hausa, + more)
- Yorùbá specialists: diacnet-yor (BiLSTM) · diacnet-yor-x (AfriBERTa) · diacnet-yor-viterbi (statistical) · diacnet-yor-db (dot-below)
- Igbo: diacnet-ig
🧭 Foundation encoder & embeddings
A Nigerian-language base model and the retrieval tools built on it.
- mist-encoder-base-ng — ModernBERT masked-LM encoder (Hausa, Yorùbá, Igbo, Nigerian Pidgin)
- naija-embed-base — cross-lingual sentence embeddings built on the encoder
🌍 Language identification (LID)
Two coverage tiers, multiple efficiency options:
- Nigerian-focused (5): lid-neural-5.1 · lid-neural-5 · lid-lite-5 (zero-dependency)
- Multilingual (25): lid-neural-25.1 · lid-neural-25.2 · lid-lite-25 (fastText)
🔤 Tokenizers
- otk-bpe-50k — 50k BPE for Nigerian languages (Yorùbá, Igbo, Hausa, Pidgin)
- otk-bpe — byte-level BPE (Swahili, Kinyarwanda, French, English)
🔎 Retrieval Stack — Search & RAG
A full embed → retrieve → rerank pipeline in compact, deployable pieces:
- naija-embed-base — embeddings
- mist-reranker-150m — ModernBERT cross-encoder for RAG
- mist-reranker-22.7M — tiny reranker for edge/CPU
⚡ Generators — Small Models, One Job Each
- mist-qg-1.5b — multilingual question generation (25+ languages, Qwen2.5-based)
- mist-tg-0.3b — 300M ByT5 title generator
🖼️ Prism — Vision Models
Compact image restoration & manipulation:
- Super-resolution: prism-upscaler-2x · prism-upscaler-4x · prism-upscaler-max
- prism-denoiser — removes noise, blur & compression artifacts
- prism-steganography — hide & recover data in images
🧠 MIST — Our LLM Line
Open text-generation models from pocket-sized to frontier-scale, with community GGUF quants available:
| Model | Size | Notes |
|---|---|---|
| MIST-Mini-8B | 8B | Compact flagship (Llama-3.1 merge) |
| MIST-Mini-8B-Thinking | 8B | Reasoning-tuned (GRPO) |
| MIST-1-70B | 70B | Mid-scale · demo Space |
| MIST-1-140B · 4-bit | 137B | Our largest release (frankenmerge) |
🗺️ What's Next
- 📊 Published benchmarks for every featured model — measured against real baselines, not claimed
🤝 Get Involved
We're a small lab with big ambitions. If you're working on African-language NLP, edtech, retrieval, or efficient specialist models — or want to use our models in production — reach out via olaverse.co.uk or open a discussion on any model.
Olaverse Lab — building AI that speaks your language.