COSMOS YZ: Cosmos T1 and Turkish Language Models 2026

COSMOSYıldız Technical University is a natural language processing research group of Computer Engineering Department and develops Turkish language models in Prof. Dr. Mehmet Fatih Amasya. The most known model of the group is Cosmos T1 (Turkish-Gemma-9b-T1), built on Gemma 2 9B, is an open weight model trained for Turkish reasoning and published on Hugging Face at 09.09.2025. The team weight in 2026 ModernBERT-TR family gave smaller and private models such as Turkish web search agent and acceleration models.

Last update: 14.10.2026

Development of Turkish language models, COSMOS AI Research GroupAmong the priorities of '. Yıldız Technical University Computer Engineering Department The group working within is sharing its models openly through Hugging Face, so Turkish artificial intelligence works offer a foundation that everyone can access.

COSMOS YZ Research Group: Academic Vision and Team

Prof. Dr. Mehmet Fatih Amasyalı, working in the leadership of COSMOS, self-star Technical University Computer Engineering Department Natural Language Processing (NLP) defines research group. The team combines academic work with engineering practice and produces model, data set and article for Turkish.

Group Members and Academic Staff:

• Prof. Dr. M. Fatih Amasyalı (Group Leader)

• Prof. Dr. Banu Diri

• Arş. See. H. Soil Section

• Arş. See. M. Kaan Supreme

• Graduated Researcher Students

2025 - 2026 Kilometer Gems and Current Developments

• Major posts from COSMOS to 2024-2026 (responsibly based on Hugging Face records and group site):

14.09.2: The Flow-Drafter draft models prepared to speed up the Qwen3.5 models were published with Apache 2.0 license (with versions tailored to the 4B, 9B, 27B).

June-July 2026: ModernBERT-TR family expanded: 150 million parameters Turkish encoder (12.06.2026) with 8.192 token context window, personal data recognition (15.06.2026), meaningal search and reorder, prompt security classifier (19.07.2026) added models.

06.03.2026: The Turkish web search is published in Turkish-Gemma-4b-T1-Scout model of the trained agent to produce evidence based on evidence.

22.02.2026: The first version of ModernBERT-TR (modernbert-tr-base-1k) is installed on Hugging Face.

09.09.2025: Cosmos T1 (Turkish-Gemma-9b-T1) is released on Hugging Face. 9 billion parameter model received 77,41 points in Turkish GSM8K results that the team share. Llama-3.1-70B-Instruct 66,13,Qwen2.5-72B-Instruct in the same table is 83,60 points.

29.05.2025: COSMOS mobile app has been published.

21.05.2025: Gemma 2 9B-based CosmosGemma (Turkish-Gemma-9b-v0.1) was announced.

13.04.2025: The "Previous Token Forecast Model" has been published to strengthen the predictive capabilities of language models.

21.09.2024: The article that tells the Cosmos-LLaVA model that can comment on the Turkish images was published in IDAP 2024 symposium; the model (Turkish-LLaVA-v0.1) was installed on Hugging Face in 31.10.2024.

Featured Academic Studies: Methodological Depth

COSMOS’s work focuses on filling gaps in Turkish AI.

1. Cosmos T1 and "Thinking" Model

Released on 09.09.2025 Cosmos T1Gemma 2 9B-based CosmosGemma model is a derived reasoning ("thinking"). It produces step-by-step thinking output before answering; adapted for mathematical, logic, multi-step out and planning tasks. 1.450 questions of the team were ranked at the first time of the table with 68,65 percent win rate in the comparison of 18 people voted. These results are based on the evaluation of the team.

2. 2. cosmosGPT: Single Language Training For Turkish

By COSMOS developed cosmosGP Home The series reveals the advantages of training large language models with only Turkish data. The study was published in INISTA conference in 2024 and examines what single language education can provide for English.

3. Create Big Web Ashliyatı and Corpus

accepted IEEE ICMI 2025 conference on 22.02.2025 "Creating a Large Clean Web Corpus for Turkish"It provides solutions to the problem of "quality data" which is the biggest obstacle in Turkish AI work. The study tells that the Turkish texts collected from the web are cleaned and transformed into a large training ash.

COSMOS models at a glance

Sectoral Impact and Future Projection u

Clear sharing of COSMOS models, an important option for institutions that care about data privacy. Since the models can be installed on their servers (on-premise), texts can be processed without being sent to a service outside. Gemma-based models are shared with the license of Gemma, the ModernBERT-TR family Apache 2.0; you need to read the relevant license conditions before using it. Other domestic models in Turkey Home Turkey's native artificial models Home You can find it in the article.

held in Istanbul in 26.07.2025 Diva: Dive into AI conferencenda COSMOS researchers presented the session titled "COSMOS YZ Research Group: Open Source English Language Model Training". As of 2026, COSMOS’s work is spread to areas such as logic, visual understanding, web search and personal data recognition besides text production.

Frequently Asked Questions

What is COSMOS YZ?

COSMOS is the natural language processing research group of Yıldız Technical University Computer Engineering Department. Prof. Dr. Mehmet Fatih produces Turkish language models, data sets and articles in the leadership of Amasya.

What is Cosmos T1, for free?

Cosmos T1 is a 9 billion parameter model trained for Gemma 2 9B based, Turkish reasoning. Weights can be downloaded free from Hugging Face; the use depends on the license conditions of Gemma. The team also opened the model.cosmosytu.com without requiring the model installation.

What model is Cosmos T1 based on?

Cosmos T1 was derived from the CosmosGemma (Turkish-Gemma-9b-v0.1) model, founded on Google Gemma 2 9B of the team.

What are the new models of COSMOS in 2026?

In 2026, Turkish-Gemma-4b-T1-Scout web search agent, ModernBERT-TR encoder family (searching, classification, personal data recognition, prompt security) and Flow-Drafter acceleration models for Qwen3.5 were published.

Are COSMOS models available in companies?

Depending on the license. ModernBERT-TR family Apache 2.0 is shared with Gemma-based models. Before corporate use, you must read the relevant license and the limits of use in the model card.