Models built for real-world language
A dedicated African-language NLP stack, tokenizers, encoders, embeddings, language identification, and diacritization, alongside merged MIST LLMs, rerankers, generators, and Prism vision models. All open-weight, all on Hugging Face.
Featured models
/ 6 MODELSA cross-section of the stack, Stable and Labs alike — identify the language, restore its tone marks, generate from it, reach for the generalist LLM, then a peek at the vision line.
Generalist
/ THE MIST LLM LINEOur merged, general-purpose chat models: Mini-8B, 70B, 140B, and reasoning-tuned Thinking variants, all built on the Llama 3.1 lineage.
MISTLABSMIST-Mini-8B
General-purpose 8B merged LLM built on the Llama 3.1 lineage with advanced model merging.
MISTLABSMIST-1-70B
General-purpose 70B merged LLM built on the Llama 3.1 lineage with advanced model merging.
MISTLABSMIST-Mini-8B-Thinking
Reasoning-tuned MIST model trained with GRPO to think step-by-step before answering.
MISTLABSMIST-1-140B
General-purpose 140B frankenmerge LLM built on the Llama 3.1 lineage with advanced model merging.
MISTLABSMIST-1-140B-4bit
4-bit NF4 quantised release of MIST-1-140B: 140B-scale weights on accessible hardware.
Specialist
/ 21 MODELSOne model, one job: DiacNet, LID, MIST Embed, MIST Rerank, MIST Gen, OTK Tokenizers, Legal Peace, and Prism — every family badge is a label here, not a page divider.
Prism is our vision family — the seed of a future OCR/vision leg, not yet part of the language stack.
DiacNetSTABLEdiacnet-1.0
Restores diacritics and tone marks across 10 languages — Yorùbá, Igbo, Hausa, Vietnamese, Polish and more. One byte-level ByT5 model, no per-language vocabulary.
MIST GenSTABLEmist-tg-0.3b
Title-generation model that names chats and documents from their content. Small, fast, and ByT5-based.
MIST EmbedLABSnaija-embed-base
Cross-lingual sentence-embedding model for Hausa, Yorùbá, and Igbo. It is the retrieval foundation for Nigerian-language search and RAG.
Legal PeaceLABSlegal-peace-v1.0
Specialised legal reasoning model fine-tuned for contract analysis and document comprehension.
DiacNetSTABLEdiacnet-1.1
Restores diacritics and tone marks across 10 languages — Yorùbá, Igbo, Hausa, Vietnamese, Polish and more. One byte-level ByT5 model, no per-language vocabulary.
MIST RerankLABSmist-reranker-22.7M
Cross-encoder reranker (22.7M-parameter micro) that re-orders retrieved passages by true relevance for search and RAG.
LIDLABSlid-neural-5.1
Neural language identification across Nigerian languages (Yorùbá, Igbo, Hausa, Pidgin, English).
LIDLABSlid-lite-25
fastText language identification across 25 world languages.
LIDSTABLElid-neural-25.1
Neural language identification across 25 world languages, tuned for long-form text: documents, articles, and passages.
MIST GenSTABLEmist-qg-1.5b
Question-generation model that turns any passage into natural questions across 25+ languages.
LIDSTABLElid-neural-25.2
Neural language identification across 25 world languages, tuned for short text: search queries, chat messages, and user input.
MIST RerankLABSmist-reranker-150m
Cross-encoder reranker (150M-parameter ModernBERT) that re-orders retrieved passages by true relevance for search and RAG.
PrismLABSprism-steganography
Neural steganography model that invisibly hides data inside images and recovers it losslessly.
MIST EmbedLABSmist-encoder-base-ng
ModernBERT masked-LM encoder pretrained for Hausa, Yorùbá, Igbo, and Nigerian Pidgin. A foundation for Nigerian-language NLP.
PrismLABSprism-upscaler-2x
Super-resolution upscaler delivering 2x image enlargement with sharp, artefact-free detail.
PrismLABSprism-upscaler-4x
Super-resolution upscaler delivering 4x image enlargement with sharp, artefact-free detail.
PrismLABSprism-upscaler-max
Super-resolution upscaler delivering maximum-quality image enlargement with sharp, artefact-free detail.
PrismLABSprism-denoiser
Image denoising model that cleans noisy photos while preserving fine detail.
DiacNetSTABLEdiactag-1.0
Restores diacritics and tone marks across 10 languages by classifying each character rather than generating text, so the output is your input with marks added and nothing else. 37.6M parameters, CPU-native.
OTK TokenizersLABSotk-bpe-50k
Byte-level BPE tokenizer built for Nigerian languages, Yorùbá, Igbo, Hausa, and Pidgin, efficient subword segmentation where general vocabularies fall apart.
OTK TokenizersLABSotk-bpe
Byte-level BPE tokenizer built for multilingual text, efficient subword segmentation where general vocabularies fall apart.
Legacy
/ SUPERSEDEDEarlier checkpoints kept on Hugging Face for reproducibility. Open weights don't expire — we just don't recommend them for new work.
| Model | Note | Superseded by | Link |
|---|---|---|---|
diacnet-yor-viterbi | Viterbi decoding, Yorùbá only | 🤗 HF | |
diacnet-yor-db | k-NN backoff, Yorùbá only | 🤗 HF | |
diacnet-ig | k-NN backoff, Igbo only | 🤗 HF | |
diacnet-yor | BiLSTM network, Yorùbá only | 🤗 HF | |
diacnet-yor-x | Transformer encoder, Yorùbá only | 🤗 HF | |
lid-neural-5 | Earlier neural checkpoint, 5 Nigerian languages | 🤗 HF | |
lid-lite-5 | Zero-dependency, 5-language logistic regression | 🤗 HF |
Build with our models
Every model here is open-weight and free to use. Fine-tune it, deploy it, or drop it straight into your own product.