Live

    Erk

    The Turkish-focused AI model that understands Turkish from the root — ranked first among open Turkish models on TurkishMMLU.

    69.7%
    TurkishMMLU
    #1 among open Turkish models
    Efficiency
    more efficient than GPT-4
    14B
    Parameters
    Turkish-focused model
    Open
    Source
    downloadable & extensible

    #1 on TurkishMMLU

    Open Turkish models — accuracy (%), identical conditions

    Erk · 14B
    %69,7
    Trendyol · 12B
    %60,9
    Turkish-Gemma · 9B
    %60,4
    Kumru · 2B
    %20,1

    Project Details

    Developed by
    eCloud Yazılım Teknolojileri
    Version
    Erk-14B · 14 billion parameters
    License
    eCloud Open Community License (open source)

    Why We Built It

    Most AI models used for Turkish today were designed for English and adapted to Turkish afterward. Every Turkish user pays the price: the model doesn't understand Turkish's agglutinative structure, over-splits words, and runs slower and more expensively. We built Erk on a strong open base model, through a Turkish training pipeline we created from scratch — with one goal: the model that understands Turkish best.

    Key Capabilities

    #1 among open Turkish models on TurkishMMLU (69.7% accuracy)

    Efficient tokenization — half the tokens of GPT-4 for the same Turkish text

    Grounded in real documents and authoritative sources; rejects false premises

    Fully open source — downloadable, inspectable, extensible

    About the Project

    Erk is a 14-billion-parameter, Turkish-focused large language model developed by eCloud Yazılım Teknolojileri. We built the Turkish training pipeline end to end — a from-scratch Turkish corpus, extensive Turkish continued pretraining, and instruction tuning grounded in real documents.

    On the independent, public TurkishMMLU benchmark, Erk ranks first among the open Turkish models tested, at 69.7% accuracy. Its strongest subjects are geography, philosophy, religion & ethics, and history. Evaluation was reported transparently, under identical conditions for all compared models.

    Erk is open source and publicly available on Hugging Face. The full training recipe is open in the nanosohbet repository. Work on the larger Erk-32B is underway.

    Technical Specifications

    14 billion parameters
    32K token context
    ChatML chat format
    Transformers-compatible
    bfloat16

    Want to Try Erk?

    Erk is open source and downloadable on Hugging Face. Contact us for commercial use and partnerships.