EVREN: Türkiye's Defense-Industry AI Platform — End-to-End Guide

EVREN is a sovereign, end-to-end AI platform developed under the Presidency of Defense Industries (SSB) of Türkiye. It brings together data labeling, active learning, distributed model training, image and large language model (LLM) inference, and model sharing under a single roof. The platform serves a broad audience — defense-industry companies, technology firms, academics, entrepreneurs, students, and developers — and runs every compute flow on bare-metal infrastructure located in Türkiye.

This guide explains, in a comprehensive and up-to-date way, what EVREN is, who can use it, its end-to-end workflow, technical infrastructure, credit economy, LLM inference API, and safe integration steps. No fake data or placeholders are used anywhere; all numeric values are taken from the platform's publicly available technical indicators.

0. TL;DR

• What: An end-to-end, contribution-based, sovereign AI platform (data → label → train → infer → share).
• Who: Developed by SSB / SSYZ (Defense-Industry AI); accessed via e-Government identity verification.
• Hardware: 64× NVIDIA H200 GPU (8.8 TB VRAM), 8× NVIDIA RTX A6000 (384 GB VRAM), InfiniBand NDR 400 Gbps.
• Models: 14 open-weight large language models, OpenAI-compatible API, streaming responses, automatic model routing.
• Economy: No direct paid subscription; earn credits (CR) by labeling data and contributing, spend them on GPU hours and inference.
• Sovereignty: LLM inference traffic is processed on the in-country bare-metal H200 cluster; data does not leave the country.
• Access: https://evren.ssyz.org.tr — free, sign in with e-Government.

1. What Is EVREN?

EVREN unifies data, AI models, and high-performance compute in a single platform. Steps that are usually managed separately — data preparation, labeling, training, testing, publishing, and operational use — are delivered end-to-end in the same environment. As a result:

• Data sovereignty is strengthened (data and models stay in Türkiye).
• Repeated infrastructure setup costs disappear.
• Community contributions are converted into credits, sharing GPU resources.
• Defense industry, academia, and the startup ecosystem meet in a common compute pool.

Core principles of the platform:

1. Sovereignty: Critical workloads run in-country without foreign-cloud dependency.
2. Contribution-based access: Instead of direct fees, data and model contributions become credits.
3. Open standards: LLM inference is exposed through an OpenAI-compatible API; existing applications integrate at low cost.
4. Scalability: Slurm-based distributed training and a reserved GPU pool scale workloads.

2. Who Builds It, Who Can Use It?

• Developer: Presidency of Defense Industries (SSB) / SSYZ.
• Contact: ssyz@ssb.gov.tr, +90 312 411 90 00.
• Address: Devlet, Süleyman Emin Caddesi 6-7, 06420 Çankaya, Ankara, Türkiye.
• Authentication: e-Government (OIDC) via kimlik.ssb.gov.tr.
• Target audience: Defense-industry firms, technology companies, academics, entrepreneurs, students, and developers.
• Language: Turkish-first interface with English information.

Access is free; there is no paid subscription model. Instead, users contribute and earn credits, then spend those credits on GPU resources.

3. End-to-End Workflow

A typical AI development cycle on EVREN consists of these steps:

1. Data upload: Images, text, or audio are transferred via presigned (pre-signed) uploads.
2. Labeling: Data is annotated with BBox, polygon, OBB, keypoint, NER, and audio-region types.
3. Pre-labeling / active learning: SAM 3 and YOLO-based models auto-pre-label; human review improves quality.
4. Training: Distributed training runs on the bare-metal HPC cluster via Slurm and Apptainer containers.
5. Evaluation: Training metrics are monitored live; models are versioned.
6. Sharing / marketplace: Trained models are shared with the community and earn credits.
7. Inference: Image inference runs on the RTX A6000 cluster; LLM inference runs on the H200 cluster through an OpenAI-compatible API.

4. Technical Infrastructure

4.1 HPC Training Cluster (H200)

This cluster is used for model training, fine-tuning, and heavy inference. Training jobs run in the Slurm queue; the credit balance is checked before a job starts.

4.2 Image Inference Cluster (RTX A6000)

This cluster targets low-latency scenarios such as real-time object detection, image analysis, and video stream processing via webcam.

4.3 Storage and Network

• Data upload is done via presigned URLs; large datasets can be uploaded in parts.
• Object storage keeps datasets and model artifacts.
• InfiniBand NDR 400 Gbps provides high bandwidth and low latency between nodes.
• NVLink NV18 accelerates intra-node GPU-to-GPU communication.

4.4 Reservation System

GPU capacity can be reserved in advance:

• Reservations can be made at earliest 24 hours and at latest 3 months ahead.
• Start and end times are set by the user.
• Cost is updated during the reservation process.
• If the balance is insufficient, the job is not started and a warning is given in advance.

4.5 Software Stack (General)

• Orchestration: Slurm (job queue, resource allocation).
• Container: Apptainer (bare-metal HPC compatible).
• Training: Distributed, multi-GPU; metrics monitored live.
• Labeling: BBox, polygon, OBB, keypoint, NER, audio region.
• Pre-labeling: SAM 3, YOLO, active learning.

5. Data Labeling

EVREN offers a broad range of labeling types for computer-vision and NLP tasks:

Active learning and pre-labeling: SAM 3 and YOLO-based automatic pre-labeling quickly marks large datasets; human review preserves quality. This flow significantly shortens labeling time.

Model marketplace and versioning: Trained models are versioned and can be shared with the community; sharing earns credits.

6. Model Training

• Training jobs run on the bare-metal HPC cluster via Slurm and Apptainer containers.
• 8 nodes × 8 H200 GPU connected over InfiniBand NDR 400G.
• Users spend GPU hours from the credit (CR) economy.
• Training metrics are monitored live.
• Cost is the product of the selected model architecture's hourly credit coefficient and the estimated training duration.
• GPU count, node count, and dataset size affect duration.
• The system checks the balance before starting; if insufficient, the job is not started and a warning is given.

7. LLM Inference

EVREN provides access to open-weight large language models running on bare-metal NVIDIA H200 infrastructure in Türkiye.

7.1 Key Features

• 14 open-weight large language models (chat, coding, reasoning, vision understanding).
• OpenAI-compatible API: existing applications connect at low integration cost.
• Streaming SSE responses.
• Automatic model routing: select by model name or use auto routing.
• Quota and credit controls: user and organization credits are separated.
• Telemetry: token, request, and fleet usage monitoring.

7.2 Data Sovereignty

LLM inference traffic is processed on EVREN's own in-country bare-metal infrastructure. API keys are scoped at account and organization level. For sensitive workloads, foreign-cloud dependency is reduced while usage and cost records are kept auditable.

7.3 Free Access Period

To broaden API adoption, API calls made until 1 November 2026 are offered unlimited and free of credit deduction. This is an opportunity for developers to integrate their existing applications with the platform and try the infrastructure.

7.4 Integration Example (General)

The following example shows secure access via environment variables (take values from your own environment; no key is written into code):

curl "$EVREN_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $EVREN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"auto","messages":[{"role":"user","content":"Hello"}],"stream":true}'

Access is also possible with Python or standard HTTP clients. Use the model name for selection, or prefer auto routing.

8. Credit Economy

EVREN operates on a contribution-based credit model rather than direct paid subscriptions.

8.1 How to Earn Credits

Credits are earned through data labeling tasks, review (validation), open dataset and model sharing, and community contributions. As the contribution score rises, the user level increases.

Contribution credit table:

Bonus credit options are also offered for certain contributions.

8.2 How Credits Are Spent

GPU training, inference, and export operations spend credits. Earned credits are used directly for costly operations such as GPU training and inference.

8.3 HOLD → SETTLE → RELEASE

The credit economy works through a three-stage chain:

1. HOLD: Credits are blocked when training starts.
2. SETTLE: Credits are finalized (spent) when the job completes.
3. RELEASE: Credits are refunded if the job fails.

This flow protects the user balance and prevents credit loss on failed jobs.

9. Platform Indicators (Current)

These figures come from the platform's public indicators and may change over time.

10. Security, Sovereignty, and Compliance

• Data sovereignty: Data and models stay on the in-country bare-metal infrastructure; LLM inference traffic does not leave the country.
• Identity: Strong authentication based on e-Government (OIDC).
• Access control: API keys are scoped at account and organization level; user and organization credits are separated.
• Auditability: Usage and cost records are kept in an auditable manner.
• Enterprise security: Enterprise security and access-control features are present in the LLM inference layer.
• Goal: Reduce foreign-cloud dependency for sensitive workloads.

11. Quick Start (General Steps)

1. Go to https://evren.ssyz.org.tr.
2. Sign in with e-Government identity verification.
3. Visit the Explore section to review datasets and models.
4. Upload a dataset or select an existing one.
5. In the labeling interface, choose the appropriate type (BBox/polygon/keypoint/NER/audio); use SAM 3 / YOLO pre-labeling.
6. Move to training: choose model architecture, GPU count, and duration; check the credit balance.
7. If needed, create a GPU reservation (earliest 24 hours, latest 3 months).
8. Try LLM inference: create an API key and send a request to the OpenAI-compatible endpoint (API calls are credit-free until 1 November 2026).
9. Share your trained model on the marketplace to earn credits.

12. Frequently Asked Questions

What is EVREN? EVREN is an end-to-end, domestic AI platform developed under the Presidency of Defense Industries. It brings together data labeling, pre-labeling (SAM 3, YOLO), distributed model training on 64× NVIDIA H200 GPU, inference, and model sharing under a single roof. The full flow is delivered with a Turkish interface and sovereign infrastructure.

Which data-labeling types does EVREN support? The platform supports bounding box (BBox), polygon, oriented bounding box (OBB), keypoint, NER text labeling, and audio-region labeling. SAM 3 and YOLO-based automatic pre-labeling significantly shorten labeling time.

How does EVREN run model training? Training jobs run on the bare-metal HPC cluster via Slurm and Apptainer containers. 8 nodes × 8 H200 GPU are connected over InfiniBand NDR 400G. Users spend GPU hours from the credit (CR) economy; training metrics are monitored live.

What is EVREN LLM inference? EVREN LLM inference is a sovereign AI service that runs chat, coding, reasoning, and vision-understanding models on bare-metal NVIDIA H200 infrastructure in Türkiye. Thanks to the OpenAI-compatible API, existing applications connect at low integration cost; requests pass through smart routing, quota, and credit controls.

Does LLM request data leave the country? No. LLM inference traffic is processed on EVREN's own in-country bare-metal infrastructure. API keys are scoped at account and organization level; user and organization credits are separated.

Which integration forms does the EVREN LLM API support? It supports the OpenAI-compatible chat-completion contract and streaming SSE responses. Developers can select directly by model name or use auto routing. Access is possible via Python SDK and standard HTTP clients.

How does the EVREN credit (CR) system work? It works through a HOLD → SETTLE → RELEASE chain. GPU training, inference, and export operations spend credits; data labeling and community contributions earn credits. Credits are blocked when training starts, finalized when the job completes, and refunded if it fails.

How are credits earned? Through data labeling tasks, review (validation), open dataset and model sharing, and community contributions. As the contribution score rises, the user level increases.

How is training cost calculated? As the product of the selected model architecture's hourly credit coefficient and the estimated training duration. GPU count, node count, and dataset size affect the duration.

13. Resources

• EVREN official site: https://evren.ssyz.org.tr
• Explore: https://evren.ssyz.org.tr/explore
• LLM Inference: https://evren.ssyz.org.tr/llm-inference
• Credit System: https://evren.ssyz.org.tr/pricing
• Infrastructure: https://evren.ssyz.org.tr/infrastructure
• SSB: https://www.ssb.gov.tr
• SSYZ: https://ssyz.org.tr

Note: EVREN is under continuous development. Numeric indicators and the model list may change over time; check the official site and documentation for current information.

This article was prepared based on EVREN's publicly available technical indicators and official statements. Access and usage conditions are determined by SSB/SSYZ.