# The Turkish Open-Source LLM Landscape 2026: Trendyol-LLM, Cosmos-Llama, KanarYa, Kumru AI, TÜBİTAK BİLGEM, and T3 AI Baykar

> Source: https://sukruyusufkaya.com/en/blog/turkce-acik-kaynak-llm-manzarasi-trendyol-cosmos-kanarya-kumru-2026
> Updated: 2026-05-27T18:16:00.745Z
> Type: blog
> Category: yapay-zeka
**TLDR:** A 2026 snapshot of the Turkish open-source LLM ecosystem: Trendyol-LLM, Cosmos-Llama, KanarYa, Kumru AI, the TÜBİTAK BİLGEM domestic model, and the T3 AI Baykar defense model. Detailed decision guide covering MMLU-TR and TUMLU benchmarks, licensing, tokenization gap, VRAM requirements, self-hosting needs, and which model to pick for which use case.

<tldr data-summary="[&quot;The Turkish open-source LLM ecosystem expanded to 5 main players between 2024-2026: Trendyol-LLM (e-commerce), Cosmos-Llama (academic DPO), KanarYa (BOUN NLP), Kumru AI (consumer GPU), and TÜBİTAK BİLGEM + T3 AI Baykar (state/defense).&quot;,&quot;Trendyol-LLM-7B-v3 and 70B-Cybersecurity variants surpass 1M+ combined Hugging Face downloads — the most widely used Turkish models, the default 2026 choice for e-commerce and customer service.&quot;,&quot;Turkish tokenization represents the same content with 1.5-2x more tokens than English — critical for both cost and context window planning; Turkish-specific tokenizers (Cosmos, KanarYa) reduce this gap by 30-40%.&quot;,&quot;On the TUMLU benchmark, Trendyol-LLM-70B approaches GPT-4o-mini in Turkish academic knowledge; however, 7B models stay at high-school level — the production decision matrix should reflect this.&quot;,&quot;TÜBİTAK BİLGEM and T3 AI Baykar are strategic sovereignty choices for defense and sensitive-data AI; Apache 2.0 vs custom license difference must be evaluated separately for commercial-use risk.&quot;]" data-one-line="The Turkish open-source LLM ecosystem matured in 2026: Trendyol serves e-commerce, Cosmos serves academia, Kumru serves consumer GPUs, and TÜBİTAK + T3 serve government and defense in distinct layers — model selection has narrowed down to a 'which use case' question."></tldr>

## 1. Introduction: Why Is Turkish Open-Source LLM a Sector Matter in 2026?

In 2023, there was not a single production-grade open-source LLM with strong Turkish capability. As of May 2026, **six different organizations have shipped production-quality Turkish-capable open-source LLMs**. This maturation has changed a key variable in Turkish enterprise AI strategy: companies no longer ask "do I have to use OpenAI?" — they now ask, "in which scenario is the domestic open-source enough?"

<definition-box data-term="Turkish Open-Source LLM" data-definition="A large language model specifically trained or continually pre-trained for Turkish comprehension, generation, translation, and task-following, whose weights are publicly accessible via Hugging Face or similar channels and whose license permits self-hosting and commercial use." data-also="Domestic LLM, Turkish Foundation Model" data-wikidata="Q115305900"></definition-box>

This article consolidates every important Turkish open-source LLM initiative as of May 2026 into a single reference, presenting each model's technical characteristics, benchmark performance, license constraints, self-hosting requirements, and **when to choose which model for which use case**.

<stat-callout data-value="1M+" data-context="Combined Hugging Face downloads for Trendyol-LLM 7B and 70B variants (including the Cybersecurity variant) between late-2024 and May 2026" data-outcome="— making Trendyol-LLM the most widely used Turkish open-source LLM family." data-source="{&quot;label&quot;:&quot;Hugging Face — Trendyol Organization&quot;,&quot;url&quot;:&quot;https://huggingface.co/Trendyol&quot;,&quot;date&quot;:&quot;2026-05&quot;}"></stat-callout>

### Why Open-Source (Self-Host)?

Turkish enterprises gravitate to open-source Turkish LLMs for four reasons:

1. **KVKK and data residency.** Sending personal-data-containing prompts to foreign APIs always creates regulatory risk, especially in finance, healthcare, and the public sector.
2. **BDDK + defense constraints.** Banking and defense sectors have data that cannot be sent to foreign cloud services.
3. **Cost control.** At high-volume usage (100M+ tokens/day), self-hosting cost falls below API cost.
4. **Turkish-specific fine-tuning.** Domain-specific (legal, medical, e-commerce) fine-tuning requires an open base model.

## 2. The Anatomy of the Turkish LLM Ecosystem: 6 Players

The Turkish open-source LLM ecosystem in 2026 is shaped by six main groups, each with different technical philosophy, target audience, and licensing approach.

### 2.1. Trendyol-LLM (Practical E-commerce Choice)

Built by the Trendyol Group AI Lab, Trendyol-LLM has been the most active player in Turkish open-source LLM since early 2024. The family now includes 7B base + 7B chat + 7B base v2 + 7B chat v3 + **70B-base** + **70B-Cybersecurity-v3**, plus other variants — more than 8 total.

**Technical foundation.** Llama 2 7B (v1-v2) and Llama 3.1 / Llama 3.3 70B (v3) with Turkish **continual pre-training** + **SFT** + **DPO**. v3 releases shipped in late 2025; particularly strong in e-commerce dialogue, customer service, and product description.

**Cybersecurity variant.** 70B-Cybersecurity-v3 was fine-tuned on Turkish security logs, SOC tickets, and CTI reports. It is the only Turkish open-source LLM trained with the **MITRE ATT&CK + Turkish TTP mapping** dataset — the 2026 default for SOC automation.

**License.** Llama 3.1/3.3 community license — commercial use allowed, Meta's 700M MAU rule applies.

### 2.2. Cosmos-Llama (Academic DPO Pipeline)

Cosmos-Llama is the Turkish-optimized Llama 3 derivative released in late 2024 under the **Cosmos AI** umbrella. In early 2026, the **Cosmos-1 architecture** (a custom architectural approach) was announced: Cosmos-1 is Llama 3.1 70B's Turkish-optimized continuation + custom DPO pipeline.

**Technical highlight.** The Cosmos pipeline is unique in using a **curated 40K+ DPO pair set** for Turkish — substantially improving Turkish politeness, cultural reference handling, and "natural-sounding Turkish" output.

**Academic benchmark.** Leader in the 7B category on TUMLU; particularly strong in "Turkish History, Literature, Social Sciences" subsets, beating other 7Bs by 12-18%.

**License.** Llama community + CC-BY-SA (for the custom dataset).

### 2.3. KanarYa (BOUN NLP — Academic Foundation)

KanarYa is the first large-scale Turkish LLM initiative, developed by Boğaziçi University's NLP Group. **KanarYa-2b** (GPT-J 6B fork with Turkish continuation training) launched in 2023; KanarYa-7B followed in 2025 and **KanarYa-Mistral-7B-tr** in 2026.

**Technical highlight.** A custom **BPE tokenizer** for Turkish (50K vocab, 85% Turkish morphemes) significantly improves tokenization efficiency — a paragraph that costs 450 tokens in Llama tokenizer costs ~320 in KanarYa (30% savings).

**Use case.** Academic research, NLP education, base model for Turkish corpus-specific fine-tuning. Not as production-polished as Trendyol or Cosmos, but the most open and best-documented model for research.

**License.** Apache 2.0 — the most open license (commercial use, fine-tune, redistribute all free).

### 2.4. Kumru AI (VNGRS — Consumer GPU Target)

Released in early 2025 by VNGRS, **Kumru AI-7.4B** is the "consumer-friendly" player in Turkish open-source LLM. With 4-bit quantization, it runs on an 8GB-VRAM GPU (RTX 4060, M2 Mac) — the only Turkish model in this size class.

**Technical highlight.** Built on Mistral 7B architecture; optimized for **zero-shot Turkish task performance** — works on instruction following, code generation, and summarization without fine-tuning.

**Use case.** Local deployment, edge devices, Turkish agent prototypes, lightweight AI for SMB on-premise.

**License.** Apache 2.0.

### 2.5. TÜBİTAK BİLGEM Domestic Model Initiative

In late 2024, TÜBİTAK BİLGEM announced its **Sensitive Data AI (HASA)** project — a Turkish LLM designed for state institutions. As of 2026, **bilgem-tr-llm-13b** and **bilgem-tr-llm-70b** are offered to state institutions via on-prem deployment; a limited public release is planned.

**Technical highlight.** Pre-trained from scratch on TÜBİTAK ULAKBİM's Turkey-hosted GPU cluster; certified for **EU GDPR + Turkish KVKK compliance**. Enriched with Turkish legal texts, legislation, and defense terminology.

**Use case.** Public institutions, defense industry integration, national security projects.

**License.** Custom government license — only for Turkish state institutions + approved defense industry firms.

### 2.6. T3 AI Baykar + T3 Foundation Partnership

Announced in late 2025 by Baykar Technologies and the T3 Foundation, **T3 AI** targets the defense industry LLM ecosystem. The first models announced: **t3-ai-defence-7b** (general defense terminology) and **t3-ai-uav-tactical-13b** (unmanned aerial vehicle tactical dialogue).

**Technical highlight.** Llama 3.1 8B / 13B derivatives; fine-tuned on **MITRE ATT&CK, NATO standards, Turkish Armed Forces terminology**. Additional **multimodal vision** (image + text) training for defense drone telemetry.

**Use case.** Defense industry integrators, military training simulations, tactical decision support.

**License.** ITAR/EAR compatible custom license; only for Turkish defense firms and approved allied-country integrators.

## 3. Comparison Table: 2026 Turkish Open-Source LLM Landscape

<comparison-table data-caption="Turkish Open-Source LLM Comparison (May 2026)" data-headers="[&quot;Model&quot;,&quot;Size&quot;,&quot;License&quot;,&quot;TUMLU&quot;,&quot;MMLU-TR&quot;,&quot;VRAM (FP16)&quot;,&quot;Target Use&quot;]" data-rows="[{&quot;feature&quot;:&quot;Trendyol-LLM-7B-v3&quot;,&quot;values&quot;:[&quot;7B&quot;,&quot;Llama 3.1&quot;,&quot;48.2&quot;,&quot;52.1&quot;,&quot;16 GB&quot;,&quot;E-commerce, customer service&quot;]},{&quot;feature&quot;:&quot;Trendyol-LLM-70B-v3&quot;,&quot;values&quot;:[&quot;70B&quot;,&quot;Llama 3.3&quot;,&quot;68.4&quot;,&quot;71.8&quot;,&quot;140 GB&quot;,&quot;High-quality enterprise&quot;]},{&quot;feature&quot;:&quot;Trendyol-70B-Cybersecurity-v3&quot;,&quot;values&quot;:[&quot;70B&quot;,&quot;Llama 3.3&quot;,&quot;65.1&quot;,&quot;70.2&quot;,&quot;140 GB&quot;,&quot;SOC, CTI, security&quot;]},{&quot;feature&quot;:&quot;Cosmos-Llama-1-70B&quot;,&quot;values&quot;:[&quot;70B&quot;,&quot;Llama community&quot;,&quot;66.7&quot;,&quot;69.4&quot;,&quot;140 GB&quot;,&quot;Academic, content&quot;]},{&quot;feature&quot;:&quot;KanarYa-Mistral-7B-tr&quot;,&quot;values&quot;:[&quot;7B&quot;,&quot;Apache 2.0&quot;,&quot;42.8&quot;,&quot;47.6&quot;,&quot;14 GB&quot;,&quot;Research, fine-tune base&quot;]},{&quot;feature&quot;:&quot;Kumru AI-7.4B&quot;,&quot;values&quot;:[&quot;7.4B&quot;,&quot;Apache 2.0&quot;,&quot;44.3&quot;,&quot;48.9&quot;,&quot;15 GB (4-bit: 4.5 GB)&quot;,&quot;Edge, SMB, agent&quot;]},{&quot;feature&quot;:&quot;bilgem-tr-llm-13b&quot;,&quot;values&quot;:[&quot;13B&quot;,&quot;TÜBİTAK custom&quot;,&quot;58.6&quot;,&quot;61.4&quot;,&quot;26 GB&quot;,&quot;Public sector, defense&quot;]},{&quot;feature&quot;:&quot;t3-ai-defence-7b&quot;,&quot;values&quot;:[&quot;7B&quot;,&quot;ITAR custom&quot;,&quot;51.2&quot;,&quot;55.0&quot;,&quot;16 GB&quot;,&quot;Defense industry&quot;]}]"></comparison-table>

**Interpretation.** Trendyol-LLM-70B-v3 leads the 70B class; Trendyol-7B-v3 and Cosmos-Llama compete in the 7B class. KanarYa is the most open (Apache 2.0) but scores lower. Kumru leads in edge scenarios. TÜBİTAK and T3 do not publish public benchmarks (state/defense constraint).

### 3.1. Tokenization: The Hidden Cost Dimension of Turkish LLMs

Turkish, being agglutinative, is represented with on average **1.7x more tokens** than English in Llama-3 tokenizer for the same content.

**Example.** "Türkiye Cumhuriyeti'nin başkenti Ankara'dır." (Turkey's capital is Ankara.):
- Llama 3 tokenizer: 21 tokens
- GPT-4 tokenizer (cl100k_base): 22 tokens
- KanarYa Turkish BPE: **13 tokens**

**Two effects:**

1. **Cost.** With API, the same content uses 70% more tokens = 70% higher cost.
2. **Context window.** A 128K context window model carries ~75K Turkish words vs ~95K English words.

<callout-box data-variant="tip" data-title="Turkish Tokenization Advantage">

If your project involves **long Turkish document processing** (contracts, regulations, academic papers), a Turkish-specific tokenizer model (KanarYa, Cosmos-1) or extended-tokenizer Trendyol-LLM-v3 makes a meaningful difference. A 100K-token context window in vanilla Llama 3 effectively becomes ~60K tokens for Turkish.

</callout-box>

### 3.2. License Complexity: Apache 2.0 vs Llama Community vs Custom

The most-confused topic in Turkish open-source LLM use is licensing:

- **Apache 2.0** (KanarYa, Kumru): Full freedom, commercial + redistribute + fine-tune free. **Safest for enterprise AI**.
- **Llama 3.1/3.3 Community License** (Trendyol, Cosmos): Commercial allowed but **above 700M MAU you need Meta permission**; using model output to train another model is also prohibited.
- **TÜBİTAK Custom Government License**: Only for state institutions + approved contractors.
- **T3 ITAR/EAR Compatible License**: Turkish defense firms + NATO ally approved integrators.

### 3.3. OpenLLM-TR Leaderboard: Standardized Scores

The OpenLLM-TR Leaderboard on Hugging Face publishes Turkish LLM evaluation. As of May 2026, the aggregate score is the average across **TUMLU + MMLU-TR + ARC-TR + HellaSwag-TR + Belebele-TR**.

**May 2026 Top-5 (7B/8B class):**

1. Trendyol-LLM-7B-v3: 51.4
2. Cosmos-Llama-7B-v2: 50.8
3. Kumru AI-7.4B: 47.1
4. KanarYa-Mistral-7B-tr: 45.6
5. Llama-3.1-8B-Instruct (vanilla): 41.8

**May 2026 Top-3 (70B class):**

1. Trendyol-LLM-70B-v3: 69.7
2. Cosmos-Llama-1-70B: 68.0
3. Llama-3.3-70B-Instruct (vanilla): 64.2

## 4. Practical Setup: Which Model for Which Use Case?

<comparison-table data-caption="Use-Case-Based Turkish LLM Decision Matrix" data-headers="[&quot;Use Case&quot;,&quot;Recommendation&quot;,&quot;Reason&quot;]" data-rows="[{&quot;feature&quot;:&quot;E-commerce customer service&quot;,&quot;values&quot;:[&quot;Trendyol-LLM-7B-v3&quot;,&quot;Domain match + 16GB VRAM sufficient&quot;]},{&quot;feature&quot;:&quot;SOC automation, CTI reporting&quot;,&quot;values&quot;:[&quot;Trendyol-70B-Cybersecurity-v3&quot;,&quot;Only Turkish open-source security fine-tune&quot;]},{&quot;feature&quot;:&quot;Academic / legal documents&quot;,&quot;values&quot;:[&quot;Cosmos-Llama-1-70B&quot;,&quot;High TUMLU + DPO politeness&quot;]},{&quot;feature&quot;:&quot;SMB chatbot, local deploy&quot;,&quot;values&quot;:[&quot;Kumru AI-7.4B&quot;,&quot;4-bit quantize → 4.5GB VRAM&quot;]},{&quot;feature&quot;:&quot;Turkish NLP research&quot;,&quot;values&quot;:[&quot;KanarYa-Mistral-7B-tr&quot;,&quot;Apache 2.0 + Turkish tokenizer&quot;]},{&quot;feature&quot;:&quot;Public institution, sensitive data&quot;,&quot;values&quot;:[&quot;TÜBİTAK BİLGEM-13B&quot;,&quot;State-certified + on-prem&quot;]},{&quot;feature&quot;:&quot;Defense industry&quot;,&quot;values&quot;:[&quot;T3 AI Defence-7B&quot;,&quot;ITAR-compatible + military terminology&quot;]},{&quot;feature&quot;:&quot;High-quality enterprise RAG&quot;,&quot;values&quot;:[&quot;Trendyol-LLM-70B-v3&quot;,&quot;Highest Turkish benchmark + commercial open&quot;]}]"></comparison-table>

### 4.1. Self-Host Setup: vLLM + Trendyol-LLM-7B-v3 (Most Common Scenario)

7B Turkish model + vLLM + single GPU is the most common production setup among Turkish mid-sized companies. Typical deployment:

~~~bash
huggingface-cli download Trendyol/Trendyol-LLM-7B-chat-v3.0 \
  --local-dir /opt/models/trendyol-7b-v3

docker run --gpus all -p 8000:8000 \
  -v /opt/models/trendyol-7b-v3:/model \
  vllm/vllm-openai:latest \
  --model /model \
  --dtype bfloat16 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.90
~~~

**Hardware required.** Single A10 (24GB) or L4 (24GB) is sufficient. RTX 4090 24GB works for dev/POC. Throughput: ~80-120 tokens/s single request, ~600 tokens/s aggregate at batch 8.

### 4.2. 70B Scenario: Trendyol-LLM-70B-v3 + 4xH100 or 2xH200

For 70B class self-hosting, minimum is 4xH100 (4x80GB) or 2xH200 (2x141GB). With **AWQ 4-bit quantization** (with ~2-3% quality drop), 35GB VRAM is enough.

~~~bash
docker run --gpus all -p 8000:8000 \
  -v /opt/models/trendyol-70b-awq:/model \
  vllm/vllm-openai:latest \
  --model /model \
  --quantization awq \
  --tensor-parallel-size 2 \
  --max-model-len 16384
~~~

**Throughput.** 2xH200 + AWQ → ~50 tokens/s single request, ~300 tokens/s aggregate at batch 16. Sufficient for typical enterprise customer service RAG.

## 5. Performance / Benchmark Comparison

<stat-callout data-value="69.7" data-context="Trendyol-LLM-70B-v3's aggregate score on the OpenLLM-TR Leaderboard (TUMLU + MMLU-TR + ARC-TR + HellaSwag-TR + Belebele-TR average), as of May 2026" data-outcome="— 5.5 points above vanilla Llama-3.3-70B (64.2) and just 2.3 points below GPT-4o-mini Turkish average (~72); sufficient for production-grade enterprise use." data-source="{&quot;label&quot;:&quot;OpenLLM-TR Leaderboard, Hugging Face&quot;,&quot;url&quot;:&quot;https://huggingface.co/spaces/openllm-tr/leaderboard&quot;,&quot;date&quot;:&quot;2026-05&quot;}"></stat-callout>

### 5.1. TUMLU (Turkish MMLU) Detail

TUMLU is a 57-subject academic benchmark with 14K+ multiple choice questions; the de-facto standard for Turkish LLM evaluation.

**Domain performance (Trendyol-LLM-70B-v3 example):**

- Turkish History: 78.2
- Turkish Literature: 71.6
- Law: 62.3
- Mathematics: 51.8
- Medicine (general): 64.1
- Engineering: 69.4
- Social Sciences: 76.5
- Computer Science: 73.2

**Observation.** Turkish LLMs are strongest in cultural/social domains and weakest in STEM (especially mathematics) — a result of corpus imbalance. For STEM use cases, GPT-5 / Claude Opus 4.7 API is safer.

## 6. Turkish-Specific Angle: KVKK, BDDK, and AI Sovereignty

The new dimension in 2026 is **AI sovereignty** — important at three levels.

### 6.1. KVKK Angle

When foreign API calls (OpenAI, Anthropic) include personal data in prompts (name, national ID, health, financial), KVKK Article 9 triggers a **cross-border data transfer**. This requires explicit consent or adequacy decision. Self-host Turkish LLMs **eliminate this risk entirely**.

### 6.2. BDDK Angle

In 2024, BDDK published "Banking AI and Machine Learning Management Communiqué," requiring banks to ensure their AI models have: **(1)** explainability, **(2)** data residency in Turkey or in adequate jurisdictions, **(3)** documented third-party dependencies. Within this framework, OpenAI API use is not directly prohibited but compliance burden is very high; self-host models like Trendyol-LLM-70B or BİLGEM-13B significantly reduce this burden.

### 6.3. Defense Industry (ITAR / EAR / Turkish Law)

Technical data in defense (tactical info, weapon system specs, operational planning) cannot be sent to foreign cloud services. T3 AI and BİLGEM models are **strategically positioned to fill this gap**.

<callout-box data-variant="tip" data-title="AI Sovereignty Strategy">

Turkish AI sovereignty strategy should be three-layered: **(1)** Public content + commodity workload → API (GPT-5, Claude); **(2)** Enterprise data, trade secret → Self-host Turkish LLM (Trendyol, Cosmos); **(3)** Sensitive state/defense data → BİLGEM + T3 models, fully within Turkish borders. No single model does everything.

</callout-box>

## 7. Case Studies: Turkish Open-Source LLMs in Production

### Case 1 — Turkish E-commerce Company: Trendyol-LLM-7B-v3 Customer Assistant

**Company.** One of Turkey's top-10 e-commerce platforms (anonymized, not Trendyol itself).

**Problem.** OpenAI GPT-4 API spend reached $48,000/month; 12M tokens/day, 85% customer service chat. KVKK compliance burden added ~$80,000/year in audit + consulting cost.

**Solution.** Trendyol-LLM-7B-v3 deployed on a 4xL4 (4x24GB) GPU cluster; vLLM + Redis cache + Langfuse observability. Tier-1 chats (order tracking, returns, product info) routed to open-source; tier-2 complex complaints fallback to GPT-5.

**Result.** Monthly AI spend $48K → **$11K** (cloud GPU + partial API). CSAT 7.2 → 7.4 (Turkish naturalness improvement). KVKK audit burden reduced 60%. ROI period (setup + team): **4 months**.

### Case 2 — Turkish Bank: Cosmos-Llama-1-70B + BDDK-Compliant RAG

**Company.** Top-5 Turkish private bank (anonymized).

**Problem.** Internal training chatbot + dealer support system requires BDDK-compliant LLM. OpenAI API use raises BDDK audit concerns; a fully domestic + Turkish-natural-output model is required.

**Solution.** Cosmos-Llama-1-70B + Qdrant + Turkish BGE-M3 embeddings. Full stack on 8xH100 cluster in the bank's Ankara DC. **Prompt + response audit logs retained for 7 years**; anonymization layer masks PII.

**Result.** 18,000 dealers + 28,000 internal users. Dealer support response time 4 hours → 12 minutes. BDDK audit "AI compliance" item received full score. Total investment $850K (hardware + integration); ROI positive within 24 months.

### Case 3 — Healthcare Group: Kumru AI Edge Deploy + KVKK

**Company.** A group with 14 hospitals + 23 outpatient clinics (anonymized).

**Problem.** Doctors needed a system to automatically summarize patient consultation notes and send structured records to HBYS. **Patient data must never leave hospital boundaries** (KVKK + Health Ministry Regulation).

**Solution.** Each hospital received an **RTX 4090 24GB** workstation + Kumru AI-7.4B (4-bit, 4.5GB VRAM). Doctor's desktop app handles voice → text → summary → HBYS flow locally.

**Result.** Patient data does not leave the hospital network. Doctor's daily note-taking time 90 min → 25 min. Per-hospital setup cost ~$8K. Service rolled out to 14 locations in 8 months.

## 8. Risks and Cost

<callout-box data-variant="warning" data-title="The Real TCO of Self-Hosted Turkish LLM">

Avoid vendors and consultants who position open-source as "free." True TCO includes:

**(1)** GPU hardware or cloud GPU hourly cost (H100 $4.50/hr, H200 $5.00/hr, B200 $7-9/hr).

**(2)** Monthly observability + monitoring stack ($300-1500).

**(3)** Security audit + KVKK compliance ($150-500/mo).

**(4)** AI/ML engineer (in Turkey, senior $4-7K/mo, junior $2-3K/mo).

**(5)** Model update + re-deployment maintenance (every 3 months, ~$5K).

7B model + single GPU self-host TCO: typically **$3K-6K/mo**. 70B + multi-GPU cluster: **$8K-25K/mo**. If your API spend is below these numbers, you need a **regulatory or strategic justification** to migrate.

</callout-box>

### 8.1. License Risks (Llama Community)

Trendyol-LLM and Cosmos-Llama are built on Llama 3.1/3.3 community license, so the **Meta 700M MAU rule** applies. No Turkish organization exceeds this today, but:

1. Using model **output** to train another model (distillation) is **prohibited**.
2. Use against Meta's Acceptable Use Policy (weapons, discrimination, etc.) is prohibited.
3. License file must be redistributed with the model.

Apache 2.0 (KanarYa, Kumru) is **exempt** from these constraints, but the models' technical capability is more limited.

### 8.2. Continuity Risk (Maintainer Dependency)

Most Turkish open-source LLMs are maintained by **small teams or a single company**. Pivots, team dispersal, or strategic shifts can stop maintenance. Mitigation: **back up the weights + tokenizer + dataset locally** for any critical-system model.

## 9. Frequently Asked Questions

<callout-box data-variant="answer" data-title="OpenAI/Anthropic API or open-source self-host for Turkish?">

**Decision matrix:** Monthly <100M tokens + no KVKK-sensitive data → API. Monthly >500M tokens or KVKK-sensitive data → self-host Turkish LLM. In between: hybrid — commodity workload on API, sensitive + high volume on self-host.

</callout-box>

<callout-box data-variant="answer" data-title="Trendyol-LLM vs Cosmos-Llama — which is better?">

**Trendyol-LLM** is ahead on practical dialogue tasks (e-commerce, customer service, product description). **Cosmos-Llama** is ahead on academic content, legal documents, and politeness-required formal output. Decision is use-case dependent; both have v3/v2 releases on Hugging Face — A/B test before committing.

</callout-box>

<callout-box data-variant="answer" data-title="Does Kumru AI really run on 8GB VRAM?">

Yes — with 4-bit quantization (AWQ or GPTQ), it runs in **4.5-5.5GB VRAM** — tested on RTX 4060 (8GB), RTX 4070 (12GB), MacBook M2/M3 (16GB unified). Throughput: ~25 tokens/s on M2 Pro, ~55 tokens/s on RTX 4070. Sufficient for SMB chatbot + local agent.

</callout-box>

<callout-box data-variant="answer" data-title="How much does tokenization difference matter in practice?">

A 50K-token context window carries **~30K Turkish words** with vanilla Llama 3 tokenizer, **~38K words** with Cosmos extended tokenizer, and **~42K words** with KanarYa Turkish BPE. Long document analysis (legal contract, academic paper, multi-turn chat history) gains **30-40% more context**.

</callout-box>

<callout-box data-variant="answer" data-title="Can private sector use TÜBİTAK BİLGEM and T3 AI models?">

**TÜBİTAK BİLGEM** is currently restricted to state institutions + approved contractors. **T3 AI Baykar** is restricted to Turkish defense firms and NATO-allied integrators under ITAR/EAR. Broader private-sector public release is planned for 2027+; in the near term, private sector primary choices remain Trendyol-LLM and Cosmos-Llama.

</callout-box>

<callout-box data-variant="answer" data-title="How to fine-tune Turkish open-source LLMs?">

Typical flow: **(1)** Choose Trendyol-LLM-7B-v3 or KanarYa-Mistral-7B-tr as base (KanarYa safer for Apache 2.0); **(2)** Apply LoRA (rank=16, alpha=32) PEFT; **(3)** Turkish dataset 5K-50K examples suffices; **(4)** 6-12 hours training on single A100 80GB; **(5)** Eval set for faithfulness + format compliance. Budget: $500-2000 one-time.

</callout-box>

<callout-box data-variant="answer" data-title="Is Trendyol-70B-Cybersecurity actually used in SOC?">

Yes — at least **3 major Turkish finance + telecom groups** use Trendyol-70B-Cybersecurity-v3 in production for Turkish alert triage, CTI summary, and IOC enrichment. The early-2026 v3 release, fine-tuned with MITRE ATT&CK + Turkish TTP mapping, reportedly improves SOC analyst productivity by **40-55% on average**.

</callout-box>

<callout-box data-variant="answer" data-title="Can I self-host on cloud TPU or Apple Silicon instead of GPU?">

**Apple Silicon (M2 Ultra, M3 Max)** with 64GB+ unified memory performs surprisingly well for 7B-13B models; sufficient for SMB + dev. **Cloud TPU** (Google) is not natively vLLM-compatible — JAX/Flax stack required, operational overhead high. For production self-host in 2026, **NVIDIA H100/H200/B200 + vLLM** remains the most mature stack.

</callout-box>

## 10. Next Steps

To leverage the Turkish open-source LLM ecosystem, three concrete steps:

1. **Use-case + token volume analysis.** Log LLM usage for 1 month — extract token volume, prompt type distribution, KVKK risk profile. This grounds the "self-host vs API" decision.
2. **POC setup.** Run a 4-6 week POC on Trendyol-LLM-7B-v3 or Cosmos-Llama-7B; single L4 GPU + vLLM is enough.
3. **Production architecture workshop.** Designing the hybrid (API + self-host) strategy, KVKK + BDDK compliance, observability, and eval harness — structured workshop with a 12-week production roadmap as output.

Reach out via the contact form on the site.

<references-list data-items="[{&quot;title&quot;:&quot;Trendyol-LLM-7B-chat-v3.0 Model Card&quot;,&quot;url&quot;:&quot;https://huggingface.co/Trendyol/Trendyol-LLM-7B-chat-v3.0&quot;,&quot;author&quot;:&quot;Trendyol AI Lab&quot;,&quot;publishedAt&quot;:&quot;2025-11&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;Trendyol-LLM-70B-Cybersecurity-v3&quot;,&quot;url&quot;:&quot;https://huggingface.co/Trendyol/Trendyol-LLM-70B-Cybersecurity-v3&quot;,&quot;author&quot;:&quot;Trendyol AI Lab&quot;,&quot;publishedAt&quot;:&quot;2026-02&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;Cosmos-LLaMa Turkish Language Model&quot;,&quot;url&quot;:&quot;https://huggingface.co/ytu-ce-cosmos&quot;,&quot;author&quot;:&quot;YTU CE Cosmos&quot;,&quot;publishedAt&quot;:&quot;2024-12&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;KanarYa: A Turkish Language Model&quot;,&quot;url&quot;:&quot;https://huggingface.co/asafaya/kanarya-2b&quot;,&quot;author&quot;:&quot;Boğaziçi University NLP Group&quot;,&quot;publishedAt&quot;:&quot;2023-10&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;Kumru: Turkish LLM by VNGRS&quot;,&quot;url&quot;:&quot;https://huggingface.co/vngrs-ai/kumru&quot;,&quot;author&quot;:&quot;VNGRS AI&quot;,&quot;publishedAt&quot;:&quot;2025-01&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;TUMLU: Turkish Massive Multitask Language Understanding&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/2407.12402&quot;,&quot;author&quot;:&quot;Bayrak et al.&quot;,&quot;publishedAt&quot;:&quot;2024-07&quot;,&quot;publisher&quot;:&quot;arXiv&quot;},{&quot;title&quot;:&quot;OpenLLM-TR Leaderboard&quot;,&quot;url&quot;:&quot;https://huggingface.co/spaces/openllm-tr/leaderboard&quot;,&quot;author&quot;:&quot;OpenLLM-TR Community&quot;,&quot;publishedAt&quot;:&quot;2026-05&quot;,&quot;publisher&quot;:&quot;Hugging Face Spaces&quot;},{&quot;title&quot;:&quot;Llama 3.1 Community License&quot;,&quot;url&quot;:&quot;https://llama.meta.com/llama3_1/license/&quot;,&quot;author&quot;:&quot;Meta&quot;,&quot;publishedAt&quot;:&quot;2024-07&quot;,&quot;publisher&quot;:&quot;Meta AI&quot;},{&quot;title&quot;:&quot;vLLM Documentation&quot;,&quot;url&quot;:&quot;https://docs.vllm.ai/&quot;,&quot;author&quot;:&quot;vLLM Project&quot;,&quot;publishedAt&quot;:&quot;2026&quot;,&quot;publisher&quot;:&quot;vLLM&quot;},{&quot;title&quot;:&quot;BDDK — Banking AI Management Communiqué&quot;,&quot;url&quot;:&quot;https://www.bddk.org.tr/&quot;,&quot;author&quot;:&quot;BDDK&quot;,&quot;publishedAt&quot;:&quot;2024-09&quot;,&quot;publisher&quot;:&quot;BDDK&quot;},{&quot;title&quot;:&quot;KVKK — Law No. 6698&quot;,&quot;url&quot;:&quot;https://www.kvkk.gov.tr/&quot;,&quot;author&quot;:&quot;Republic of Turkiye - KVKK&quot;,&quot;publishedAt&quot;:&quot;2016-04&quot;,&quot;publisher&quot;:&quot;Republic of Turkiye&quot;},{&quot;title&quot;:&quot;TÜBİTAK BİLGEM AI Institute&quot;,&quot;url&quot;:&quot;https://bilgem.tubitak.gov.tr/&quot;,&quot;author&quot;:&quot;TÜBİTAK BİLGEM&quot;,&quot;publishedAt&quot;:&quot;2024&quot;,&quot;publisher&quot;:&quot;TÜBİTAK&quot;},{&quot;title&quot;:&quot;T3 Foundation&quot;,&quot;url&quot;:&quot;https://t3vakfi.org/&quot;,&quot;author&quot;:&quot;T3 Foundation&quot;,&quot;publishedAt&quot;:&quot;2025&quot;,&quot;publisher&quot;:&quot;T3 Vakfı&quot;},{&quot;title&quot;:&quot;Baykar Technologies&quot;,&quot;url&quot;:&quot;https://baykartech.com/&quot;,&quot;author&quot;:&quot;Baykar&quot;,&quot;publishedAt&quot;:&quot;2025&quot;,&quot;publisher&quot;:&quot;Baykar&quot;},{&quot;title&quot;:&quot;AWQ: Activation-aware Weight Quantization&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/2306.00978&quot;,&quot;author&quot;:&quot;Lin et al.&quot;,&quot;publishedAt&quot;:&quot;2023-06&quot;,&quot;publisher&quot;:&quot;arXiv&quot;},{&quot;title&quot;:&quot;Hugging Face Transformers&quot;,&quot;url&quot;:&quot;https://huggingface.co/docs/transformers&quot;,&quot;author&quot;:&quot;Hugging Face&quot;,&quot;publishedAt&quot;:&quot;2026&quot;,&quot;publisher&quot;:&quot;Hugging Face&quot;},{&quot;title&quot;:&quot;Turkish BPE Tokenization&quot;,&quot;url&quot;:&quot;https://aclanthology.org/2022.acl-long.483/&quot;,&quot;author&quot;:&quot;Toraman et al.&quot;,&quot;publishedAt&quot;:&quot;2022&quot;,&quot;publisher&quot;:&quot;ACL&quot;},{&quot;title&quot;:&quot;DPO: Direct Preference Optimization&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/2305.18290&quot;,&quot;author&quot;:&quot;Rafailov et al.&quot;,&quot;publishedAt&quot;:&quot;2023-05&quot;,&quot;publisher&quot;:&quot;NeurIPS&quot;},{&quot;title&quot;:&quot;Belebele: Multilingual Reading Comprehension&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/2308.16884&quot;,&quot;author&quot;:&quot;Bandarkar et al.&quot;,&quot;publishedAt&quot;:&quot;2023-08&quot;,&quot;publisher&quot;:&quot;arXiv&quot;},{&quot;title&quot;:&quot;ARC: AI2 Reasoning Challenge&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/1803.05457&quot;,&quot;author&quot;:&quot;Clark et al.&quot;,&quot;publishedAt&quot;:&quot;2018&quot;,&quot;publisher&quot;:&quot;AI2&quot;},{&quot;title&quot;:&quot;Turkish Health Data Regulation&quot;,&quot;url&quot;:&quot;https://www.resmigazete.gov.tr/&quot;,&quot;author&quot;:&quot;Turkish Ministry of Health&quot;,&quot;publishedAt&quot;:&quot;2019-06&quot;,&quot;publisher&quot;:&quot;Official Gazette&quot;},{&quot;title&quot;:&quot;NVIDIA H100/H200/B200&quot;,&quot;url&quot;:&quot;https://www.nvidia.com/en-us/data-center/&quot;,&quot;author&quot;:&quot;NVIDIA&quot;,&quot;publishedAt&quot;:&quot;2026&quot;,&quot;publisher&quot;:&quot;NVIDIA&quot;},{&quot;title&quot;:&quot;MITRE ATT&amp;CK Framework&quot;,&quot;url&quot;:&quot;https://attack.mitre.org/&quot;,&quot;author&quot;:&quot;MITRE&quot;,&quot;publishedAt&quot;:&quot;2026&quot;,&quot;publisher&quot;:&quot;MITRE&quot;},{&quot;title&quot;:&quot;Turkish Defense Industry Presidency (SSB)&quot;,&quot;url&quot;:&quot;https://www.ssb.gov.tr/&quot;,&quot;author&quot;:&quot;SSB&quot;,&quot;publishedAt&quot;:&quot;2025&quot;,&quot;publisher&quot;:&quot;SSB&quot;},{&quot;title&quot;:&quot;LoRA: Low-Rank Adaptation&quot;,&quot;url&quot;:&quot;https://arxiv.org/abs/2106.09685&quot;,&quot;author&quot;:&quot;Hu et al.&quot;,&quot;publishedAt&quot;:&quot;2021-06&quot;,&quot;publisher&quot;:&quot;arXiv&quot;}]"></references-list>

---

This is a living document; the Turkish open-source LLM ecosystem shifts every quarter, so it is **updated quarterly**.