Turkey's AI Platform EVREN Opens to Users: 64 Nvidia H200 GPUs and Open Models
Originally published in Turkish. Read the original →
EVREN, the national AI platform developed by the SSB, is now available with a high-performance Nvidia H200 GPU cluster, a pool of open-weight LLMs and an architecture built around data sovereignty.
At a time when AI technologies have become a geopolitical source of power on a global scale, data sovereignty and local compute capacity form the foundation of countries' digital independence. The Presidency of Defence Industries of Turkey (SSB) has officially opened access to EVREN, the national platform it developed to meet Turkey's infrastructure needs in AI.
Embracing a broad audience that ranges from the defence industry ecosystem to universities, and from tech startups to individual researchers and developers, EVREN brings high-performance computing infrastructure together with a community-based collaboration model.
Turkey's Data Sovereignty Move: What Is the EVREN Platform?
EVREN is an integrated platform that aims to meet, with local resources, the compute and data processing needs that are the biggest bottleneck for developers and institutions working in AI. The platform's core vision is to guarantee that data produced in Turkey is processed within the country's borders, while running modern open-weight AI models on a national cloud backbone.
Instead of imposing a paid subscription up front, the platform is built around a model in which participants earn compute credits in proportion to what they contribute to it, and it aims to speed up academic research and commercial ventures.
Supercomputing Power: Bare-Metal HPC Architecture with 64 Nvidia H200s
At the compute core of the platform sits a high-performance Bare-Metal HPC (High Performance Computing) cluster designed for training modern deep learning and large language models. The hardware is configured to run industry-standard AI workloads at maximum efficiency:
- 64 Nvidia H200 GPUs: A total of 8.8 TB of high-bandwidth VRAM.
- 1,536 CPU Cores and 24 TB of DDR5 RAM: Infrastructure optimized for preprocessing massive datasets and in-memory data management.
- 25.6 Tb/s InfiniBand Backbone: A 400 Gbps InfiniBand NDR network providing ultra-low-latency communication between nodes.
- 47.5 Thousand TFLOPS of Compute Performance: High processing capacity for multi-node training and complex inference architectures.
Users can use these resources for model fine-tuning, training models from scratch and large-scale inference tasks. The hourly GPU cost ranges between 10 and 35 credits depending on the workload, and the platform also has a flexible reservation mechanism. Developers can reserve training runs as early as 24 hours and as late as 3 months ahead, guaranteeing their compute time in advance.
Real-Time Computer Vision: 8 Nvidia RTX A6000 Server
EVREN is not limited to text-based language models; it also hosts a dedicated image inference server for those working on image and video analytics:
- 8 Nvidia RTX A6000 GPUs: A total of 384 GB of VRAM and 10,752 CUDA cores.
- 50 Millisecond Inference Latency: Support for real-time object detection on webcam, drone or fixed camera streams.
- More Than 40 Concurrent Sessions: Resource sharing optimized for vision models running in parallel.
In the image processing layer, SAM-2 (Segment Anything Model 2) and current YOLO architectures are offered as direct integrations. Users can upload raw images to the platform and prepare datasets with bounding box (BBOX), polygon, keypoint and text labeling tools, and then train their models directly in the same environment.
14 Open-Weight Large Language Models and an OpenAI-Compatible API Ecosystem
EVREN's language model infrastructure runs 14 of the most capable large language models shared worldwide as open source and open weights on the local HPC cluster in Turkey. The platform's notable architectural features are:
- OpenAI-Compatible API Standard: Developers can connect their applications to the EVREN infrastructure in seconds by changing only the base URL and API key in their existing code base.
- Dynamic Model Routing and Streaming Support: Real-time token streaming is available for chat, coding, reasoning and multimodal tasks.
- Unlimited and Free Access Until November 1, 2026: To test the infrastructure and speed up integrations, all API calls made until November 1, 2026 are offered without limits and without being deducted from the credit balance.
Contribution-Based Credit System and Community Ecosystem
EVREN's sustainability formula rests directly on community contributions. Institutions and researchers using the platform can earn GPU credits by uploading datasets, labeling data or sharing the models they train with the community:
- 100 Data Units: 2 Credits
- 500 Data Units: 5 Credits
- 1,000 Data Units: 15 Credits
- 5,000 Data Units: 25 Credits
- 10,000 Data Units: 40 Credits
Bonus credits are also available for special tasks and verified model contributions. As of launch, the platform has 11.2 thousand registered users, 1.1 thousand datasets, 631 active models, 14.4 million data items and 82.7 million verified labels.
How to Access the EVREN Platform and What Is on the Roadmap?
The platform's future vision includes integrating domestic and national large language, vision and speech models into the system, and pooling the hardware power of different institutions into a single pool through distributed GPU management.
All Turkish citizen researchers, students and engineers who want to join EVREN and start using its compute resources can easily access the system at evren.ssyz.org.tr via e-Devlet (Turkey's government e-services portal) identity verification.
