Every Turkish-tuned open-weight model we’ve verified on Hugging Face. Operator-curated catalog — specs from published model cards, no vendor marketing copy. Updated as new releases ship from Trendyol, YTU CE COSMOS, VNGRS, and the broader Turkish AI community.
English-trained LLMs handle Turkish poorly because of how BPE tokenization interacts with Turkish’s agglutinative morphology. A single Turkish word (e.g., evlerimizdekilerden) can fragment into 8+ tokens on a Llama 3 tokenizer, vs ~3 on Qwen 3 (which trained for 119 languages) and ~1 on a purpose-built Turkish tokenizer. More tokens means more compute, more VRAM, and degraded attention quality on the same logical sentence.
That’s why a Turkish-from-scratch model (Kanarya), or a continued-pretrained Turkish fine-tune (the YTU CE COSMOS family, Trendyol’s flagship), often outperforms a larger English-base model on Turkish-language tasks despite having a fraction of the parameters. The Turkish-language community has its own ecosystem, and the catalog below covers the active releases.
Each entry links to the model page on /models with the same depth of editorial we give every model. Where we’ve run benchmarks (TurkishMMLU specifically), the score appears next to the model. Where we haven’t yet, it says “not yet tested” — we don’t fake numbers.
31B-parameter Turkish-tuned reasoning model with i1-imatrix quantizations by mradermacher. Designed for step-by-step problem solving in Turkish. Highest download count among large Turkish-tuned models.
35B MoE (3B active) tuned specifically for Turkish financial-services text — bank statements, investment research, accounting terminology. Niche-cluster model from the Turkish fintech community.
Turkish-from-scratch language model trained by Ali Safaya (Koç University researcher). Named after the kanarya (Turkish for 'canary'). Trained on 250+ GB of Turkish text including Wikipedia, news, and books.
Turkish BART-style sequence-to-sequence model fine-tuned specifically for summarization. Not a chat model — purpose-built for input-document → Turkish-summary pipelines.
Smaller Kanarya variant — 750M parameters. Runs on CPU or 4GB GPU comfortably. Useful for low-resource Turkish text classification, embeddings, or completion tasks where latency matters more than quality.
GPT-2 Large architecture trained from scratch on Turkish. Reference baseline for measuring how much modern instruction-tuned models actually improve on the GPT-2 era.
Kumru 2B is a compact Turkish text-generation model from VNGRS. The Hugging Face config reports a Mistral-family architecture with an 8K context window, and the public Ollama build makes it a practical edge-speed Turkish
Turkcell LLM 7B v1 is an Apache-2.0 Turkish text-generation model built on a Mistral architecture. The measured Ollama artifact uses a RefinedNeuro GGUF distribution of the public Turkcell model.
Mistral 7B v0.2 continued-pretrained on Turkish data + instruction-tuned. The 32K context window makes it the best Turkish open-weight model for long-document work today.
Mistral Turkish v2 is a public Ollama-distributed Turkish Mistral variant. The upstream Hugging Face repository was not publicly accessible during intake, so commercial-use status is kept conservative until the source li
Malhajar Mistral 7B Turkish is an Apache-2.0 Mistral 7B Instruct v0.2 Turkish fine-tune. The benchmarked Ollama tag is a koezgen quantized distribution of the public malhajar model.
YTU's Turkish-tuned Gemma 2 9B model. The highest community-rated Turkish-language LLM on Hugging Face by likes-to-downloads ratio as of May 2026. Continued pre-training on Turkish corpora followed by instruction tuning.
Trendyol LLM Asure 12B is a Gemma 3 based multimodal instruct model for Turkish and English business workflows. The public Ollama build used in local testing is the alibayram GGUF distribution.
Gemma 4 26B MoE (4B active params) pruned and Turkish-tuned. The largest Turkish-tuned open-weight model on HF as of May 2026. MoE architecture means it loads 26B of weights but runs at 4B-active speed.
YTU Turkish Gemma 9B v0.1 is a Gemma 2 based Turkish instruction model from the YTU CE COSMOS ecosystem. The benchmarked Ollama tag is an alibayram GGUF distribution of that public model.
Llama 3 8B continued pre-trained on Turkish corpora, then instruction-tuned for Turkish chat. YTU CE COSMOS group's most-downloaded Llama variant. GGUF builds available — drops into Ollama directly.
Turkish-tuned chat model released by Trendyol, Turkey's largest e-commerce platform. Built on Llama 2 7B, fine-tuned on Turkish customer-service style dialogues plus general Turkish chat. The first major Turkish LLM from
YTU CE COSMOS's Llama 3 8B Turkish instruction-tuned variant. Follow-up to the original Turkish-Llama-8b that uses the Llama 3 base instead of Llama 2 — better tokenizer, stronger zero-shot Turkish.
Base (non-chat) variant of Trendyol's 7B Turkish LLM. The chat sibling is the more popular pick; this base version is for operators building their own instruction-tuning pipeline on top of Trendyol's pre-training.
TurkishMMLU is a 900-question, 9-subject Turkish-language benchmark for general knowledge and reasoning. Scores appear next to each model as our compute finishes. Each result links to the raw test-run Gist for verification — same trust apparatus as our /benchmarks/coding (HumanEval+) leaderboard.
Every catalog entry above passes a discovery gate (downloads ≥ 50 on Hugging Face, updated within the last 18 months, actually Turkish — not multilingual with a Turkish token in the slug). If a model meets those criteria and isn’t listed, point us to the HF repo and we’ll add it on the next sweep.