Turkish Language Models in 2026: A Verified Landscape

Search for “Turkish open-source LLM” and some of the top results list models that do not exist. Invented repository names, invented benchmark scores, and a US export-control regulation presented as a model licence. This article does the opposite: every model name links to a real Hugging Face repository, every number says where it came from, and nothing I could not verify appears here at all. At the end there is a script that lets you check any list yourself in twenty seconds.

A warning first: models that do not exist#

A list page currently ranking near the top of Turkish search results contains these model names: KanarYa-Mistral-7B-tr, Kumru AI-7.4B, Cosmos-Llama-1-70B, TÜBİTAK BİLGEM-13B, bilgem-tr-llm-70b, and a t3-ai-defence-7b said to be published under an “ITAR/EAR Custom License”.

None of those repositories exist — and ITAR/EAR is US export-control law, not a model licence.

Do not take my word for it either. Verification takes thirty seconds, and this script runs:

from huggingface_hub import HfApi
from huggingface_hub.utils import RepositoryNotFoundError

api = HfApi()
candidates = [
    # claimed, unverifiable repositories
    "asafaya/KanarYa-Mistral-7B-tr",
    "vngrs-ai/Kumru-AI-7.4B",
    "ytu-ce-cosmos/Cosmos-Llama-1-70B",
    # ones that genuinely exist
    "asafaya/kanarya-2b",
    "vngrs-ai/Kumru-2B",
    "ytu-ce-cosmos/Turkish-Llama-8b-v0.1",
]

for repo in candidates:
    try:
        info = api.model_info(repo)
        print("EXISTS ", repo, "| license:", info.card_data.get("license"))
    except RepositoryNotFoundError:
        print("MISSING", repo)

The rule to apply when reading any model list: if a row gives no repository ID and no licence, it is a claim, not information. The Turkish AI ecosystem’s biggest problem is not a shortage of models; it is a shortage of verifiable writing about them.

The verified landscape#

Every row below was checked against Hugging Face while writing.

The most striking thing about this table is what is missing from it: there is no Turkish-specific open-weight model above 12B. The ecosystem’s real centre of mass sits between 2B and 12B, and it is surprisingly strong on the encoder side (BERT and ModernBERT derivatives).

That has a direct consequence. If you need a large model for Turkish, the answer is not a Turkish-specific model but a multilingual one that handles Turkish well. The cleanest-licensed option in that category is utter-project/EuroLLM-22B-Instruct-2512 — 22.6B parameters, Apache-2.0, with Turkish explicitly among its 35 supported languages.

Licences: what “open” does and does not mean#

“Open source” is the single most misapplied label in these listings. The verified position:

• Turkish-Llama-8b-v0.1 is under the Llama 3 licence. Not Apache — it carries use restrictions and attribution obligations.
• Turkish-Gemma-9b-v0.1 and Trendyol-LLM-Asure-12B are under the Gemma licence. Also not Apache-2.0 — labelling these two as “Apache” is the most common error in Turkish listings.
• EuroLLM-22B-Instruct-2512 genuinely is Apache-2.0.

BILGE and T3: the state-backed side#

BİLGE is the national language model developed by TÜBİTAK BİLGEM. What is verifiable: it was trained on more than one trillion tokens, scored 44.44 out of 50 on a cultural-translation evaluation, and has an official presentation site. Two things are not verifiable, and saying so is part of the job: the parameter count is contested — different outlets describe a family spanning 1.9B to 122B versus a single 37B model — and no source confirms that BİLGE is open-weight. Calling it “Turkey’s open-source model” is therefore not supportable today.

T3 AI, developed by the T3 Foundation with Baykar and launched at TEKNOFEST, has a beta available. Claims circulating about an export-controlled defence variant have no primary source behind them.

The benchmark reality: language-specific does not mean better#

This is the most overlooked finding in the whole discussion. Cetvel, from Koç University’s KUIS AI group (arXiv:2508.16431, accepted to EACL 2026), evaluated 33 open-weight models up to 70B across 23 tasks and found that Turkish-centric instruction-tuned models generally underperform relative to multilingual or general-purpose models, despite being tailored for the language. Llama 3 and Mistral beat Turkish-specific models on many tasks.

That is not an argument against Turkish models. It is an argument that a language label is not a performance guarantee. A 9B model with continued Turkish pretraining may well beat a multilingual model of the same size on Turkish — but there is no guarantee it beats a strong 70B multilingual one.

Where to look on the benchmark side: TurkBench (arXiv:2601.07020, 8,151 samples, 21 subtasks, EACL 2026 SIGTURK), the older TR-MMLU reference (arXiv:2501.00593), and for live standings yerel.ai/leaderboard plus the alibayram TR-MMLU table on Hugging Face. Self-reported model-card scores are usable too — the Turkish-Gemma-9b-v0.1 card reports a 63.31 average on Turkish benchmarks — but they are the publisher’s own measurement, not independent, and should be labelled that way.

A practical selection flow#

1. Is the task classification, tagging or search? You do not need a generative model. The ModernBERT-TR family (~0.1B) is dramatically faster and cheaper for this.
2. Doing semantic search? Pick a dedicated Turkish embedding model and encode both queries and documents with the same one.
3. Generative work where data cannot leave the organisation? The 8B–12B continued-pretraining models are a reasonable starting point; read the licence first.
4. Need the quality ceiling? Stop looking for a Turkish-specific model and benchmark a strong multilingual one on your own task set.
5. In every case: decide from a 50–200 example evaluation set built from your own data, not from a leaderboard. Public leaderboards do not measure your job.

What the Turkish AI ecosystem needs is not more lists but verifiable ones. For the wider open-weight picture see open-source LLMs, for how to measure a model on your own data see LLM evaluation, and for why a small model is often enough see small language models.