PRODUCTION MODELS · OPEN WEIGHTS

    Models built for real-world language

    The models we run in production, led by PurpleMIST, our frontier decision models, alongside language identification across 608 languages, diacritic restoration in 11, and question generation in 25.

    Browse the models 🤗 View on Hugging Face
    quickstart.py
    from transformers import pipeline
    # Which of 608 languages is this?
    lid = pipeline("text-classification", model="olaverse/lid-neural-608")
    lid("Ẹ kú àárọ̀, ṣé dáadáa ni?")
    {'label': 'yor_Latn', 'score': 0.99999...}
    Apache-2.0
    LICENCE
    608
    LANGUAGES IDENTIFIED
    11
    LANGUAGES DIACRITIZED
    25
    LANGUAGES FOR QUESTION GEN

    Production models

    / READY FOR PRODUCTION

    Grouped by family: each family is one model line, offered in the sizes we run in production. PurpleMIST, our frontier decision line, comes first; the specialists identify the language, restore its marks, and generate questions from it.

    FrontierOur flagship models, going head to head with models hundreds of times their size.

    PurpleMIST

    FRONTIER/ Decision models

    Our frontier models. Give them a message, ticket or record and any number of typed questions, and they return calibrated probabilities for every answer in one pass, in English and African languages.

    PurpleMIST

    PurpleMIST-Flash-1.0

    A 7.9B decision model that matches DeepSeek-V4-Pro’s accuracy on LocalLLaMA/typed-decisions in English (0.720, zero-shot) and on African Typed Decisions across 8 languages (0.668), at about 1/200 of the size.

    56 1
    View
    PurpleMIST

    PurpleMIST-Mini-1.0

    The same decision interface in 1.9B parameters, for English workloads. Zero-shot on LocalLLaMA/typed-decisions it scores 0.601, ahead of decision models up to four times its size.

    49 0
    View
    SpecialistSmall models that do one job extremely well, built to run cheaply in production.

    LID-608

    / Language identification

    One model line in two sizes, sharing 608 languages, the same labels and the same two modes. Neural for accuracy on short text, lite for raw speed on a CPU.

    LID-608

    lid-neural-608

    Identifies 608 languages, built African-first. The most accurate model in our tests, and 0.906 on 1-5 word inputs in traffic mode, ahead of GlotLID and Meta’s fastText LID.

    29 0
    View
    LID-608

    lid-lite-608

    The same 608 languages, labels and modes in a 37 MB fastText model. About 6,800 texts per second on a single CPU thread, for bulk filtering and edge deployments.

    0 0
    View

    DiacNet

    / Diacritic restoration

    Three ways to put the marks back in 11 languages, Arabic included: two generative models at two sizes, and a character tagger that can never change your letters.

    DiacNet

    diacnet-2.0

    Restores diacritics, tone marks and Arabic vowel marks in 11 languages from one 582M byte-level model. The most accurate Olaverse diacritizer on 8 of 10 languages, with meaning hints and optional typo correction.

    521 0
    View
    DiacNet

    diacnet-mini-2.0

    The same 11 languages and features as diacnet-2.0 at half the size. Close to the full model on accuracy, cheaper to serve.

    481 0
    View
    DiacNet

    diactag-2.0

    Restores marks in 11 languages, Arabic included, by tagging each character instead of generating text, so it can never change your letters. 37.9M parameters with a 38.6 MB int8 ONNX export for CPU.

    23 0
    View

    MIST QG

    / Question generation

    The question-generation line of MIST: a compact model that turns passages into the questions people would ask, in 25 languages. More MIST lines will sit alongside it.

    MIST QG

    mist-qg-1.5b

    Turns any passage into natural, search-style questions across 25 languages. Built for edtech, FAQs, and minting retrieval training data.

    406 0
    View

    Which one?

    / PICK BY CONSTRAINT

    Each job has a large and a light option. Pick on accuracy, hardware, and whether the model may touch your letters.

    If you needUseWhy
    Calibrated decisions on messages and tickets, African languages includedFrontier: matches DeepSeek-V4-Pro accuracy at 7.9B: 0.720 on LocalLLaMA/typed-decisions (English, zero-shot) and 0.668 on African Typed Decisions (8 languages).
    English decisions on smaller hardwareSame interface at 1.9B: 0.601 on LocalLLaMA/typed-decisions (English, zero-shot).
    Best diacritic accuracy, Yorùbá and European languagesLowest error on 8 of 10 languages (mean DER 0.0117). Meaning hints and optional typo correction.
    Generative diacritics on a smaller budgetSame features at 300M, half the size, mean DER 0.0141.
    Diacritics that can never change a letter, on CPUCharacter tagger: compliance 1.0000 by construction. 38.6 MB int8 ONNX, about 8 ms a sentence on a laptop CPU. Strongest on Igbo, Vietnamese and Modern Standard Arabic.
    Language ID on user input and short textTraffic mode: 0.960 web-weighted accuracy, 0.906 on 1-5 word inputs.
    Language ID at volume, CPU-only or at the edge37 MB, about 6,800 texts/s on one CPU thread. Same 608 labels and modes.
    Questions from passages, or retrieval training pairs25 languages, 0.955 round-trip keep-rate on a held-out eval.

    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.

    View on Hugging FaceFollow on GitHub