Turkish Artificial Intelligence Models: BİLGE, Kumru, T3 AI and Others

The large language models have a very basic fact that they are reviewing while crocheting the world: No matter how talented giant models with silicone valley outlets, the work becomes different when it comes to English. Most of us have noticed the strange translation cochan phrases that are put together when working with these models, meaningless of idioms, or simply extending a simple technical problem to unnecessary place. Here, at this point, domestic artificial intelligence projects have become a technical obligation, not only being a national pride.

Today, the artificial intelligence ecosystem developed in Turkey is not a simple interface in front of OpenAI or Anthropic models. A strong infrastructure is built with the contribution of open-source communities trained with its own weights and focused directly on the state-of-the-art structure of Turkish.

How to Denersin?

A brief table first: how you can access this model today and whether the model is switched on (14.10.2026).

Hugging FaceDownloadable

Open, non-commercial academic use alone m

Home

Home

Private to institutions, platform built into institution

Close

Why Global Models Are Not Yet?

While global AIs are trained with trilyons, a little part of this data consists of Turkish. This situation brings two major issues:

Token Waste and High Billing:

token systems designed according to English cannot solve the root and additional logic of Turkish. A simple word divides 3-4 pieces (token). This adds both API usage bills and extends the system’s response generation time in bar.

Cultural and Legal Context:

It is quite risky to take a contract analysis suitable to the Turkish Commercial Code of a global model or to classify the customer’s complaint with the right feeling. Because the model references Western culture and American law logic as default.

In order to close this deficit, the academy focused on developing Turkish artificial intelligence models by combining public and independent initiatives in Turkey.

What It Works On The Site? Featured Native Models

Artificial intelligence projects in Turkey Home; responds to different needs from state-supported large-scale infrastructures to agile solutions.

1. BİLGE: Extensive Language Model Family from TÜBİTAK

The most concrete and comprehensive step on the public side is implemented by TÜBİTAK BİLGEM News The project was. The biggest claim of BİLGE is to use Turkish as a native language without translating from another language.

Different Sizes For Different Needs:

BİLGE is not a single model, a family. With 1B designed for mobile and end devices, the performance and efficiency balance were trained from 9B to scratch. BİLGE 27B has been developed with Continuous Pre-Training method with Turkish data. Designed for complex reasoning tasks, BİLGE 122B is still in the development phase.

Devasa English Korpus:

The training started with about 1 trillion vocabulary raw Turkish data pool. 250 billion high quality tokens were selected from this pool and the data was enriched with 500 billion synthetic data. The aim is to grasp the model’s idioms, appliques and technical terms in Turkish context.

Low Energy & Cost:

It is aimed to spend less processing power because it uses a token that is appropriate to Turkish. Public institutions are designed for corporate applications where data security is critical.

Access status: As of 14.10.2026, BİLGE is not open to general use. For pilot projects according to the official site, TUBITAK TICKET can be contacted. Access-related announcements were announced only to be made from the official channels of TÜBİTAK with the Ministry of Industry and Technology.

2. 2. Kumru AI: Fast, Focused and Lightweight

In the world of venture, models focus on functionality rather than entering the huge parameter race. From Kumru AI, A Turkish LLM alternative that businesses can run on their servers without spending huge budgets.

Text abstracting, categorizing customer demands and establishing natural Turkish dialogue Very successful modelprovides high speed of distractions without removing hardware resources.

VNGRS trains from scratch for Turkish Kumru 7,4 billion parameters and can work on a GPU with 16 GB VRAM. Home You can try the demo on it; 2 billion parameters Kumru-2B is open with Apache 2.0 license on Hugging Face. For technical details and limits Kumru AI guide News See.

3. T3 AI: Open Source and Community Power

T3 AI developed in Baykar business association with Turkish Technology Team Foundation (T3 Foundation) represents the community size of the work. In Model 11.07.2025, TEKNOFEST has been opened to use with beta version in the Social Platform. During the announcement, it was announced to contribute to the project of 1,792 volunteers from 67 provinces.

The project says it was developed as open source; but the number of parameters and model weights were not yet shared where to be downloaded. If the weights are opened, it can be easy to train special models in areas such as law or medicine.

4. Academic Spine and Outdoor Weight Initiatives

The ecosystem in Turkey is not only limited to three names:

Trendyol LLM:

The first version of Trendyol LLM (2024) was trained with 10 billion tokens on the LLaMA 2 7B model of Meta. Next versions passed Qwen 2.5 7B, Qwen 3 8B and Gemma 3 12B bases. Trendyol-LLM-Asure-12B shared in February 2026 can also work and set for e-commerce tasks, such as abstracting, question-cevap. Models are mainly shared on Hugging Face.

TOURNA:

TURNA, developed by TABILab in Boğaziçi University Computer Engineering Department, is a 1,1 billion parameter encoder based on UL2 frame, a Turkish language model. Designed for both understanding and text generation tasks and shared for non-commercial academic research on Hugging Face.

Zemberek Heritage:

Open source English natural language processing library HomeOne of the tools that have been applied for many years in basic works such as root-ex solutioning.

5. MAIN GPT (HAVELSAN): Native Brain of Defense and Corporate Safety

One of the most critical feet of domestic AI moves is thrown on the defense industry and public security side. Developed to protect in-house data security and high privacy standards by HAVELSAN HAVELSAN MAIN GPT It stands out as one of the corporate AI infrastructures that can work off Turkey.

The root (Multi Artificial Intelligence Network) promoted in 01.02.2024 of HAVELSAN, is a platform where institutions can establish artificial intelligence assistant based on their data. The closed circuit is designed to work; the purpose does not leave the data outside the institution. At the first stage, text-based works; content abstracting is able to process information from open source data and code writing.

6. Aigency: Early Term Turkish Language Model Initiative Home

Şanlıurfa-based eCloud Software Technologies started developing in 2022 Aigencyintroduces itself as one of the first domestic language models in Turkey. According to the news on the local press, Meta’s LLaMA is based on 3 models and focuses on Turkish text production, content analysis and corporate communication processes. An independent assessment of the model has not yet been published.

Overview of models:

Low server cost, quick response time

T3 AI

T3 Foundation and Baykar

Parameter count not disclosed

Community Based Turkish LLM

Volunteer data set, open source ecosystem target

Trendyol LLM

Trendyol Tech

7B - 12B outdoor weight

E-Commerce and Customer Dialogue

Directly integrated into practical scenarios

HOME GPT

Home

Corporate & Isolated Language Model

Defense, Public and Corporate Data Security

Closed circuit (on-premise) operation

Aigency

eCloud Software

Early Stage Language Architecture

Text Production and Corporate Communication

Early term initiative focusing on Turkish text production

Why Important for Developers and Companies?

Choosing a native language model is not only the idea of "we support the ground technology"; directly hosting the advantages of companies interested in their pocket and legal processes:

KVKK and Data Security:

A serious headache due to the regulation of a bank, hospital or e-commerce site to receive user data and send it to an API abroad. Local models do not exit the data from the institution because they can work directly on their server (on-premise).

Less Hardware, More Speed:

Instead of hundreds of billion parameters bulk models, you can easily run on a single middle segment GPU, a 7B or 14B compact model trained for a specific job. That’s why server and electricity costs are seriously down.

Break Out Dependency:

Global artificial intelligence companies can double API prices at a night or bring access restriction to your business to stop. Outdoor weights Hugging Face native models that you can download from platforms like, allow you to keep your software architecture completely in your own control.

With the opening of national supercomputer resources such as TRUBA and carrying out software experience acquired in the defense industry (ASELSAN, HAVELSAN), artificial intelligence production in Turkey is rapidly spreading to the base. Now the matter is not only to print text; to create autonomous systems that really understand our native culture and language. Native models are also fully pushing the spine of this independence.

Writing the Future With Our Own Code

Artificial intelligence race is now measured not only with who has a larger server farm, but how realistic it touches the local needs of this technology. The models developed in Turkey; instead of copying the ready-made molds, it is placed in a line that presses many more feet, focusing on solving the subtleties of the Turkish, providing data security and reducing operational costs of companies. This point, from the corporate power of TICGE, Kumru AI and T3 AI, shows that an independent technology infrastructure that thinks and produces in its own language is no longer a appeal, but a tangible reality. To build the future with our own codes instead of consuming with others’ algorithms, the most critical step we can jump in digital age.

Some of the companies behind these models Turkish AI companies that stand out in 2026 Home We got in the article. For the technical details of Kumru Kumru AI guide look at people working on Turkish language models Masqot communities you can meet.