This training is designed for AI engineers, ML engineers, data scientists, and platform engineers who want to run open-source large language models with high quality on Turkish tasks and bring them into a KVKK-compliant self-hosted infrastructure for regulated sectors that require data privacy. At the heart of the program is the following approach: productizing open-source LLMs is not simply downloading a model to a server and running it. Real enterprise value comes from selecting the right base model (DeepSeek V3 / R1, Qwen 3, Gemma 3, Llama 3.3), deciding on fine-tuning by Turkish task type, choosing correctly among Ollama / vLLM / TGI as the inference engine, building a Turkish RAG with an embedding and vector-DB stack, optimizing hardware cost via a quantization strategy, establishing on-prem deployment in line with KVKK 'cross-border transfer' rules, and binding all of this to an auditable governance layer.
DeepSeek fundamentally transformed the open-source LLM ecosystem during late 2024 and throughout 2025. The DeepSeek V3 (671B total parameters, 37B active parameters in a Mixture-of-Experts architecture), V3.1 (hybrid reasoning), and R1 (open-source reasoning model) series offered open-source alternatives to closed reasoning models like OpenAI o1 and Claude Opus 4.7 Deep Think. Distill models (from R1-Distill-Qwen-1.5B to 70B) enabled smaller companies and solo developers to access reasoning capabilities at low cost. This training covers the DeepSeek ecosystem with architectural depth: MoE mechanics, FP8 native quantization, V3.1 hybrid reasoning mode, the R1 chain-of-thought training paradigm, and the correct use-case scenarios for distill models are covered in detail.
The training's Turkish focus is another critical dimension. Participants gain the competence to survey, compare, and correctly choose open-source base models with strong Turkish performance. The Qwen 3 (Alibaba) family offers a broad range from 0.5B to 72B with strong Turkish coverage; the Gemma 3 (Google) open-weight family is strong with multimodal capabilities; Llama 3.3 (Meta) 70B is consistent in Turkish instruction following. Alongside these, Turkish-fine-tuned local models — Trendyol LLM (e-commerce-focused), KUIS Cosmos LLM (Koç University general-purpose), AYDA, the BERTurk family — hold significant positions in their niches. The training teaches systematically evaluating these models with MTEB Turkish, Belebele, MMLU-TR, TruthfulQA-TR, and organization-specific custom eval sets.
The training covers local execution and production-inference layers together. On a developer machine and a single-node server, how DeepSeek-R1-Distill, Qwen 3, and Turkish models can be run in seconds with Ollama and LM Studio; GGUF format management, quantization-level selection, and exposing OpenAI-compatible APIs are shown hands-on. On the production side, vLLM with PagedAttention and continuous batching, Text Generation Inference (TGI) with Hugging Face native deployment, and SGLang with structured generation and constrained decoding are covered comprehensively. Tensor parallelism and multi-GPU deployment address high-throughput needs.
Perhaps one of the strongest modules of the program is dedicated to Turkish fine-tuning with LoRA and QLoRA. Adapting open-source base models to organization-specific Turkish tasks saves up to 99% of hardware costs compared to full fine-tune via the PEFT (Parameter-Efficient Fine-Tuning) approach. The training covers end-to-end Turkish instruction-dataset preparation (Alpaca, ShareGPT, ChatML formats), using SFTTrainer with the Hugging Face TRL library, hyperparameter selection (learning rate, batch size, gradient accumulation), adapter merging, GGUF conversion, and final model deployment. As a result, participants reach a level where they can train a Turkish-optimized custom LLM for their own company.
Turkish RAG architecture is also one of the program's core modules. To strengthen open-source LLMs with Turkish document-based systems, the comparison of embedding models like multilingual-e5-large, jina-embeddings-v3, and bge-m3; Turkish-morphology-aware chunking strategies (recursive, semantic, sentence-window); self-hosted vector-DB deployment with Qdrant, Weaviate, and pgvector; BM25 + vector hybrid search architecture; and a cross-encoder reranker layer with bge-reranker-v2 are covered in detail. This provides an architectural foundation directly applicable to enterprise knowledge management, customer service, document summarization, and compliance products.
Quantization strategies are a critical component of the program. GGUF (Q2_K, Q4_K_M, Q5_K_M, Q6_K, Q8_0), AWQ, GPTQ, FP8 native, EXL2, and other modern quantization methods are addressed comparatively; measuring quantization-induced quality regression, selecting levels by hardware constraints, and the self-hosted cost model (GPU, electricity, ops) are covered in detail. Thus, participants reach a level where they can make the right quantization choice from an architectural perspective in the triangle of quality, speed, and cost.
A distinguishing point of the program is a module dedicated to KVKK-compliant on-prem and air-gapped deployment. In banking (BDDK), energy (EPDK), healthcare (SGK), and regulation-heavy sectors, self-hosted LLMs are not merely a technological preference but a mandatory architectural decision in terms of data privacy, regulation, and audit requirements. The training comprehensively covers model download and signature verification, the transfer process in restricted-network environments, on-prem Kubernetes clusters and the GPU operator, network segmentation and mTLS, PII masking and audit-log layers, governance documentation, and a compliance checklist.
In the capstone project, each participant designs an end-to-end Turkish LLM stack tailored to their own organization: use case → base model → fine-tune decision → inference engine → embedding and vector DB → KVKK-compliant deployment topology. Participants present this architecture together with a diagram, deployment plan, and eval report, receiving peer review and instructor feedback. By the end of the training, participants will have the technical and architectural competence to understand the DeepSeek and Turkish open-source LLM ecosystem at a strategic level, professionally establish local and production-inference layers, perform Turkish LoRA fine-tuning, design a Turkish RAG architecture, optimize their quantization strategy, and build a KVKK-compliant self-hosted AI infrastructure. The training consists of 3 days, 12 modules, and over 70 hands-on lessons.