Kumru AI What is? Native Turkish Artificial Intelligence Model

Kumru is a large language model (LLM) with 7,4 billion parameters that VNGRS trains from scratch for Turkish. Text abstracting was developed for jobs such as document processing and question-cevap; the first general version was announced at the end of September 2025. The architecture is based on Mistral-v0.3; the rock window (sliding window) is equivalent to LLaMA-3 as the structure is closed.

How to Denersin Kumru?

1. From Browser Home Go to the address; the demo chat screen of the model with 7,4 billion parameters opens.

2. Write the text you want to sum up Turkish issues or text to the chat box.

3. 2 billion parameters if you want to try on their computer or server Kumru- 2BYou can download ' from Hugging Face; open with Apache 2.0 license.

Adjust the expectations accordingly: Kumru is a small model designed to be run in the institution, not a general purpose chat assistant. The demo version can give false information and be forced in math; we tell the reasons below.

Kumru AI developed by VNGRS company was designed specifically to respond to corporate needs that require data privacy and high performance. Although this development brings the discussions that the ChatGPT rival would not be, the main mission of Kumru is from the last user, to provide solutions to the regulatory sectors in Turkey (on-premise).

What is Kumru AI and Why Developed from Zero?

Kumru AI is a solvent (decoder-only) architecture based on 7.4 billion parameters Big Language Model (LLM). The main motivation in the development of the model is the deep need for a basic LLM model, which can be used in the organization, considering the data security regulations of Turkey and excellence in the Turkish language.

• Zero Education Advantage: Instead of making fine adjustment on multilingual models such as LLaMA or Gemma, Kumru was a strategic decision to train from scratch for English. Ready-made models are spent much more tokens when processing tokens (text encoders), which increases the cost of training and use. Kumru uses a modern tokenizer that is specially developed to Turkish. According to the comparison of VNGRS, other multilingual models spend more than 38% more tokens than Kumru for the same Turkish text. This efficiency makes Kumru ideal for distribution in terms of speed and cost.

• Lightweight and Accessibility: Kumru can work on medium-level GPUs such as RTX A4000 or RTX 3090 with 16 GB VRAM, it can be integrated into enterprise systems without requiring large hardware investments. This feature is subject to data security regulation sectors (finance, health, etc.) critical importance.

Kumru How Artificial Intelligence Was Educated and What Are Technical Specifications?

Development process of KumruBy a small team consisting of Turkish engineers, it was carried out with a careful engineering work:

• Preliminary Training Stage: The model was cleaned 500 GB of size using NVIDIA H100 and H200 GPUs in the pre-training phase of 45 days. Kumru in this process was exposed to 300 billion tokens.

• Architecture and Support: The architecture of Kumru is based on Mistral-v0.3; equivalent to LLaMA-3 as structure as the rock window is closed. In training decisions such as bulk business size, optimized and learning speed Technical report of LLaMA-3 u base was taken.

• Context Length: Kumru 8.192 has context length of Turkish token. This means you can document up to about 20 A4 pages in one transaction.

Answers to questions from VNGRS: How to Position Native Artificial Intelligence Kumru?

Following the intense interest shown in Kumru, VNGRS clears the positioning of the model by answering the most critical questions from social media (X):

Kumru Competite to ChatGPT?

Answer: No. Kumru is a small, fast and cheap solution developed for in-house (B2B) solutions, not a product for the end user (B2C). ChatGPT is a huge model family positioned for both B2C and B2B.

Why Do I Do Errors (Relation)?

Answer: The problem of incorrect information in artificial intelligence models (halfination) is minimized at the stage of consolidated learning (RLHF). Kumru, like the first Davinci models of OpenAI, it can provide incorrect information as it does not yet pass through the RLHF stage. This deficiency will be resolved with future improvements.

Why is Kumru Math?

Answer: Language models are composed of words and they learn statistically from the text they train information; they do not have a calculator mechanism in their own. Kumru is developed for Turkish competencies and is now mistaken in math because it does not pass through the RLHF stage.

Why Choose Zero Education?

Answer: Prepared (pre-trained) models (such as LLaMA) and require fine adjustment, use the token of the original model (text encoder). This increases the cost up to double while processing Turkish texts. VNGRS has chosen to train Kumru with a modern token that has been optimized in terms of special, cost and speed to the needs of Turkish from completely scratch. This provides efficiency and cost advantage in the long run.

7 Billion Model Will Be Published Open Source?

Answer: As larger and multi-mode models are produced, it is planned to be released by old models starting from 7 billion version. 2 billion parameters alone on VNGRS's Hugging Face page as of 14.10.2026 Kumru- 2B and the basic version of Kumru-2B-Base is open with Apache 2.0 license. 7,4 billion parameter model has not yet been published as open source.

The Letters in the Word I mentioned are Wrong Answering, Why?

Answer: The smallest unit for popular language models is word or word (token). The model sees tokens, not the letters that make it when processing texts. Therefore, they often make mistakes in questions about letters.

Mistral What Does Model Architecture Come?

Answer: The model architecture Mistral does not mean that a leading trained Mistral model has been taken and trained. Kumru was created and trained from end to end with tokenizer and model weights. Architectural defines the overall structure of the model, such as rough architecture of the building.

Kumru How to Offer AI Use and What is the Future Plans?

Kumru AI, sandru.ai address via a demo version with access.

Current status (14.10.2026): Since the first release in September 2025, VNGRS's Hugging Face page has not been released a new Kumru model. The following plans aim that VNGRS shared with the first version.

Future Plans:

• Cheerful Learning (RLHF): It is planned to quickly take this stage to minimize the errors of the model and improve the mathematical ability.

• Multimodality: It is aimed to create a multimodal model of Kumru by providing visuals and scanned documents processing ability (image encoder integration).

• Larger Models: The larger and ruined Kumru versions are planned to be trained with larger and up-to-date data sets and then broadcast as open source of old models.

These developments, VNGRSIt also shows the mission of making long-term investments to Turkey’s own language private AI ecosystem.

Other language models developed in Turkey (BİLGE, T3 AI, Trendyol LLM, TURNA and MAIN) Turkey's native AI models Home We compare in the article. VNGRS Turkish AI companies that stand out in 2026 Home You can also browse the list. If you want to talk on Turkish language models Masqot communities You can find what suits you.