Türkiye's defence industry AI platform EVREN opens with 11 open-weight language models
Anadolu Agency reported on September 27, 2026, citing a statement from the Presidency of Defence Industries, that EVREN, the AI platform developed within the Presidency, has opened for use. An inference layer covering 11 open-weight large language models is now live, existing applications can connect through an API, and API calls made until November 1, 2026 will not be deducted from credit balances.
Data, models and compute on one platform behind an e-Devlet sign-in
The Anadolu Agency report, filed from Ankara, is based on a statement from Türkiye's Presidency of Defence Industries (SSB). According to the report, EVREN was opened for use to strengthen the country's technological independence and data sovereignty. It is aimed at a wide audience, from defence companies and technology firms to academics, entrepreneurs and students, and it brings data, AI models and high-performance compute together on a single platform. Access runs through evren.ssyz.org.tr with identity verification via e-Devlet, the Turkish government's online services gateway.
The same statement was also carried by TRT Haber, DefenceTurk and BirGün, with TRT Haber and BirGün both crediting Anadolu Agency as their source. DefenceTurk also added screenshots of the platform's own pages to its report. EVREN itself is not a new name. A June 19, 2026 notice from Ankara University's Artificial Intelligence Institute had already reported that EVREN, an end-to-end MLOps platform developed by SSB, was open for use by undergraduate and graduate students. The e-Devlet sign-in and the contribution-based credit model were there at that point too, and promotional starter credits carried a deadline of July 31, 2026. That June announcement listed data labelling, model training, GPU support and testing as its scope; the large language model inference layer appears in the September report.
Contribution instead of payment, plus an 11-model inference layer
Per the statement relayed by Anadolu Agency, EVREN does not sell access directly; it runs on credits earned through contribution. Users collect credits by uploading datasets, labelling data or sharing models they have trained, then spend those credits on model training and large language model inference in the platform's high-performance GPU pool. Screenshots of the platform that DefenceTurk added to its report put numbers on this: contribution tiers run from 2 credits for 100 data items up to 40 credits for 10,000, labelled imports earn a 3-credit bonus, and GPU time is charged on a tiered scale of 10 to 35 credits per GPU per hour. For computer vision work such as object detection, segmentation and classification, the platform offers an end-to-end path from data labelling through to model training.
The technical weight of the announcement sits with the language models. An inference layer covering 11 open-weight large language models has been switched on, and thanks to its API architecture, real-time response support and automatic model routing, existing applications can plug into the EVREN infrastructure. A DefenceTurk screenshot shows that models for tasks ranging from chat and coding to reasoning and visual understanding are called through a single OpenAI-compatible API. To broaden use, API calls made until November 1, 2026 will be offered without limit and without drawing down users' credit balances. Two user figures are in circulation: Anadolu Agency says the platform quickly reached 7,500 active users, while TRT Haber's version of the same agency text speaks of more than 8,000 active users and says in its headline that the platform passed 8,000. The user counter on the platform screen shared by DefenceTurk also reads 8.0K. Neither the Anadolu Agency nor the TRT Haber text states the date at which its figure was measured.
Data stays in the country, and the API is credit-free until November 1
The statement puts data sovereignty up front. According to Anadolu Agency, prompts, responses and usage data on the platform are processed on high-performance GPU infrastructure located in Türkiye, which keeps sensitive data from being sent to cloud services abroad. For businesses and developers that are cautious about moving data out of the country, that makes it a new option for large language model experiments. The credit-free API period until November 1, 2026 also opens a limited window for testing chatbot or agent prototypes at low cost.
The statement sketches a longer horizon as well. The plan is to extend the platform with domestic and national large language, image and speech models, and to strengthen its compute through a distributed GPU management approach. The stated aim is a shared, Türkiye-based AI development infrastructure for the defence industry, universities, technology companies, entrepreneurs and developers. It is still early for a production decision, though: the sources do not include the names of the 11 models, the terms for commercial use, or how many credits API calls will cost after November 1.
The story is from Türkiye itself, and our commentary is marked separately
Here the facts about Türkiye come straight from the sources. EVREN is presented as a national AI platform developed within the Presidency of Defence Industries and is used through evren.ssyz.org.tr with e-Devlet identity verification. According to Anadolu Agency, prompts, responses and usage data are processed on GPU infrastructure in Türkiye, and API calls until November 1, 2026 are not deducted from credit balances. The active user count is 7,500 in the Anadolu Agency report and more than 8,000 in TRT Haber's. EVREN is intended to contribute to work under the Artificial Intelligence Action Plan.
What follows is UNALSOFT commentary, not sourced fact. There are points a business should settle before building EVREN into its own product, and none of them appear in the sources: the API documentation and how to obtain keys, how long data is retained and whether it is used for model training, and whether foreign users without an e-Devlet account can get access. The decision should therefore wait until the platform's own terms have been read. We covered a separate story on access to compute, the GPU for Everyone program, on September 1; that article is independent of this announcement.
UNALSOFT's take
For businesses, the practical meaning is this: an infrastructure where prompts and responses are processed in the country and several open-weight models are reachable through one API becomes one of the options that teams avoiding sending data abroad can weigh for chatbot and agent experiments. The period until November 1, 2026 is a sensible test window to connect an existing workflow and measure answer quality with your own scenarios, without using real customer data. Before moving to production, the model names, the terms of use and how many credits API calls will consume after November 1 need to be clear. That is why in Agentic AI projects we design the model layer to be swappable from the start; changing the infrastructure should not mean rewriting the workflow.
Have you decided which infrastructure your AI pilot should run on?
A short conversation is enough to settle model choice, data flow and a test plan together.