Turkey

AI models, vendors, and hardware buying realities from Turkey. Curated by myaihardware.com.

Models

Vendors

Languages

Why Turkey matters for AI builders

If you're building AI hardware today, Turkey offers a surprising edge: NVIDIA H100 and A100 supply is often available here for immediate purchase through regional distributors when other markets face months-long backlogs, a critical advantage for scaling proof-of-concept clusters. The dominant model families driving demand in Turkey are Turkish-optimized large language models like Trendyol's variants for e-commerce and custom whisper-sized speech models for the country's booming call-center automation sector. The underrated angle is Istanbul's role as a cross-continental data-flow hub — its terrestrial cable links to Europe and Asia make it a low-latency deployment zone for edge inference serving both the Middle East and Eastern Europe simultaneously.

Market context

Turkish agglutinative morphology breaks standard BPE tokenizers, inflating token counts significantly compared to English. Academic groups at Bogazici, METU, ITU, and Yildiz Teknik (ytu-ce-cosmos) lead open-source Turkish LLM efforts. Trendyol leads the commercial space with Trendyol-LLM-7b and the newer Trendyol-LLM-Asure-12B, while Turkcell, Hepsiburada, and others maintain internal models with minimal public release. State-backed TUBITAK BILGEM has announced TR-LLM initiatives, but public weights are not yet released as of 2026.

Models (25 of 25)

Filter by vendor, license, or parameter range.

LicenseVendorParameter range

Turkcell-LLM

unverified-public-info · Turkey

Restricted

Params: 13B
Context: ,
Min VRAM: ,
GGUF: No

Internal commercial model for Turkcell, details unverified.

Trendyol-LLM-Asure-12B

Trendyol · Turkey

Open

Params: 12B
Context: 32,768
Min VRAM: 9 GB
GGUF: Yes

Trendyol's current flagship, Mistral-Nemo based, with strong Turkish reasoning and 32K context.

Kocdigital-LLM-8b-v0.1-Q4_K_M-GGUF

KocDigital (Koc Holding) · Turkey

Open

Params: 8B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Pre-quantized Q4_K_M GGUF of Koc Holding's Turkish Llama-3 fine-tune; runs in ~6GB VRAM.

Kocdigital-LLM-8b-v0.1

KocDigital (Koc Holding) · Turkey

Open

Params: 8B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Enterprise-ready Turkish Llama 3 fine-tune.

Turkish-Llama-8b-v0.1

Yildiz Technical University (ytu-ce-cosmos) · Turkey

Open

Params: 8B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Continued-pretrained Llama-3 base on large Turkish corpus; widely used as a base for further fine-tuning.

Turkish-Llama-8b-Instruct-v0.1-GGUF

Yildiz Technical University (ytu-ce-cosmos) · Turkey

Open

Params: 8B
Context: 8,192
Min VRAM: 5 GB
GGUF: Yes

Pre-quantized GGUF variants (Q4_K_M, Q5_K_M, Q8) for direct llama.cpp / Ollama use on Turkish local builds.

Turkish-Llama-8b-Instruct-v0.1

Yıldız Teknik Üniversitesi · Turkey

Open

Params: 8B
Context: 8,192
Min VRAM: 6 GB
GGUF: Yes

Turkish instruction-tuned Llama 3, fluent generation.

Cosmos GPT

unverified-public-info · Turkey

Restricted

Params: 7B
Context: ,
Min VRAM: ,
GGUF: No

Publicly unconfirmed Turkish LLM from cosmos.com.tr.

Hepsiburada AI

unverified-public-info · Turkey

Restricted

Params: 7B
Context: ,
Min VRAM: ,
GGUF: No

E-commerce focused Turkish LLM, unverified.

LlamaTurk-7b

METU NLP Group · Turkey

Open

Params: 7B
Context: 4,096
Min VRAM: 5 GB
GGUF: No

Systematic METU study on continued-pretraining vs vocabulary-extension vs LoRA for Turkish; multiple checkpoint variants released.

Trendyol-LLM-7B-Base

Trendyol · Turkey

Open

Params: 7B
Context: 4,096
Min VRAM: 5 GB
GGUF: Yes

Strong Turkish language base model with good fluency and domain coverage for e-commerce.

Trendyol-LLM-7B-Chat

Trendyol · Turkey

Open

Params: 7B
Context: 4,096
Min VRAM: 5 GB
GGUF: Yes

Fine-tuned on high-quality Turkish instruction data, very strong in Turkish conversational and e-commerce tasks.

Trendyol-LLM-7b-chat-v1.0

Trendyol · Turkey

Open

Params: 7B
Context: 8,192
Min VRAM: 5 GB
GGUF: Yes

Second-generation Trendyol chat model based on Mistral-7B with improved Turkish reasoning and instruction following.

TR-LLM

unverified-public-info · Turkey

Restricted

Params: 7B
Context: ,
Min VRAM: ,
GGUF: No

State-funded Turkish LLM project, details unverified.

Kizagan-E4B-Turkish-Reasoning

Alican Kiraz (community) · Turkey

Open

Params: 4B
Context: 8,192
Min VRAM: 4 GB
GGUF: Yes

Small Turkish reasoning fine-tune of Gemma-3-E4B with explicit chain-of-thought training for Turkish queries.

Kanarya-2B

Ali Safaya (KUIS AI Lab, Koc University) · Turkey

Open

Params: 2B
Context: 4,096
Min VRAM: 2 GB
GGUF: Yes

Lightweight model trained from scratch on Turkish corpora, suitable for low-resource Turkish NLP.

turkish-gpt2-large

Yıldız Teknik Üniversitesi · Turkey

Open

Params: 774M
Context: 1,024
Min VRAM: 3 GB
GGUF: No

Largest Turkish GPT-2, strong text generation.

Kanarya-750M

Ali Safaya (KUIS AI Lab, Koc University) · Turkey

Open

Params: 750M
Context: 1,024
Min VRAM: 2 GB
GGUF: No

Smallest viable open Turkish base LM; useful for embeddings, classification heads, and resource-constrained fine-tuning.

mt5-base-turkish-summarization

Mukayese (Turkish NLP benchmark consortium) · Turkey

Open

Params: 580M
Context: 1,024
Min VRAM: 2 GB
GGUF: No

Released alongside the Mukayese Turkish NLP benchmark - the standard reference for Turkish summarization eval.

VBART-Large

VNGRS (vngrs-ai) · Turkey

Open

Params: 387M
Context: 1,024
Min VRAM: 2 GB
GGUF: No

First Turkish-only BART pretrained from scratch; specialized variants for summarization, QA-generation, paraphrasing, title generation.

turkish-gpt2-medium

Yildiz Technical University (ytu-ce-cosmos) · Turkey

Open

Params: 355M
Context: 1,024
Min VRAM: 1.5 GB
GGUF: No

Compact Turkish-only generator suitable for classification heads, autocompletion, and embedding bases.

turkish-gpt2

Yıldız Teknik Üniversitesi · Turkey

Open

Params: 124M
Context: 1,024
Min VRAM: 1 GB
GGUF: No

Lightweight Turkish GPT-2, easy to fine-tune.

bert-base-turkish-cased

Bavarian State Library (dbmdz) · Turkey

Open

Params: 110M
Context: 512
Min VRAM: 1 GB
GGUF: No

De-facto standard Turkish BERT used in nearly every Turkish NLP pipeline since 2020; massive community fine-tune base.

bert-base-turkish-sentiment-cased

Savas Yildirim (academic) · Turkey

Open

Params: 110M
Context: 512
Min VRAM: 1 GB
GGUF: No

Production-ready Turkish sentiment classifier; minimal-setup inference for product reviews and social media.

turkish-base-bert-uncased

Yildiz Technical University (ytu-ce-cosmos) · Turkey

Open

Params: 110M
Context: 512
Min VRAM: 1 GB
GGUF: No

Workhorse Turkish encoder; widely used for NER, classification, semantic search in Turkish production systems.

Hardware buying realities

Country-grade truths on customs, currency, and where to actually source GPUs.

The Turkish Lira's precipitous depreciation against the US dollar and Euro makes hardware procurement for AI builders a nightmare of unpredictable pricing. A GPU that costs $1,600 at launch can swing wildly in TL terms within weeks; retailers reprice daily or even hourly on platforms like Hepsiburada and Trendyol, often pegging listings to a shadow dollar rate well above the official Central Bank rate. This volatility forces builders to either hoard cash in foreign currency or buy immediately before the lira weakens further, turning a technical purchasing decision into a currency speculation game. For instance, an RTX 4090 that was 45,000 TL in early 2024 could exceed 80,000 TL by year-end purely due to FX pass-through, even if global MSRPs remain static.

Notable vendors

TrendyolKoç University (KUIS AI Lab)KoçDigital (Koç Holding)Yıldız Technical University (ytu-ce-cosmos)METU NLP GroupVNGRSMukayese consortiumTÜBİTAK BİLGEMTurkcellHepsiburadaCosmos GPT

Run benchmarks

llama.cpp, MLPerf, and inference leaderboards.

VRAM calculator

Will this model fit on your GPU?

Hardware guides

Local LLM setup, homelab, mini-PC builds.

Methodology

How we benchmark, verify ASINs, and source data.

Beginner's guide

New to local AI? Start here.

Sourcing & commerce disclosure

Model metadata on this page is compiled from each vendor's public HuggingFace repo, model card, or research paper at time of indexing. Performance figures (KMMLU, MMLU, HAERAE, AraBench, IndicGEM, etc.) link directly to source benchmarks where available; missing or partial values are shown as ", ".

Outbound hardware links (GPUs, mini-PCs, accelerators) are Amazon Associate links using store ID fredoline-20. As an Amazon Associate we earn from qualifying purchases. Prices and availability shown elsewhere on the site are snapshots, not live quotes — verify on Amazon before purchasing. See About / disclosures for the full policy.