All insights
ORBITRA INSIGHTS

The Turkish AI & LLM Landscape in 2026: Who Is Building What

A fact-checked map of Türkiye's LLMs and AI programmes in 2026: BİLGE, T3 AI, HAVELSAN MAIN, Kumru, Turkcell, Trendyol, the 2026–2030 action plan, and what to build on.

By Orbitra AIPublished Updated 10 min read
AI LandscapeTurkeyLLMAI Automation
The Turkish AI & LLM Landscape in 2026: Who Is Building What

If you run an organisation in Türkiye and you are weighing an AI project, one question quietly shapes every other decision: whose model do you build on? For most of the last few years, the honest answer was "an American one, in English, adapted for Turkish as best we can." That is no longer the only answer. Türkiye now has state-backed model programmes, telecom and retail companies adapting models to Turkish, a from-scratch Turkish model family, and, since August 2026, a national AI action plan with concrete targets. This article maps who is building what, and what that means when you have a real project to ship.

We have kept this grounded. Where a fact is public and verifiable, we state it. Where a claim is announced but not yet delivered, we say so. This edition was fact-checked and updated in October 2026; model versions change quickly, so treat specific numbers as a snapshot, not gospel.

Why a "Turkish LLM" matters at all

A language model trained overwhelmingly on English text can speak Turkish, but it does so with a foreign accent you can feel in production. Turkish is agglutinative—meaning is stacked onto word roots through long chains of suffixes—so a single dictionary word can carry what English needs a whole clause to express. Models that never saw enough Turkish tokenise it inefficiently, misread morphology, and stumble on the cultural and legal context that business answers depend on.

For a consumer chatbot, that shows up as slightly-off phrasing. For an enterprise workflow—an assistant answering questions about a KVKK policy, a bank summarising a Turkish contract, a support agent handling Istanbul slang—it shows up as wrong answers delivered confidently. That gap is what the Turkish-model ecosystem exists to close, and why it is worth understanding before you commit a budget.

The national and state-backed models

The most striking development is that Turkish LLMs are now treated as public infrastructure, not just commercial products. The state-backed efforts, however, are at very different stages of availability.

BİLGE is the research-council entry, developed by TÜBİTAK BİLGEM and unveiled in June 2026. Its official site lists a family of models: 1B and 9B versions trained from scratch, a 27B version built by continued pre-training of existing open-source models, and a 122B version still in training. As of October 2026 the site states that BİLGE has not yet been made available for general use; in July 2026 the industry and technology minister said it would open to developers soon. Treat it as one to watch, not one to build on today.

T3 AI ("T3 AI'LE") was developed by the Turkish Technology Team (T3) Foundation with defence company Baykar, with contributions from about 1,800 volunteers across 67 provinces and partners including the Ministry of National Education, SETA, TRT, the Turkish Academy of Sciences, Anadolu Agency, and Microsoft. A public beta ran briefly on the TEKNOFEST social platform in July 2025, and model weights were shared with hackathon participants, but we found no public release of the weights. It signals sovereign intent more than it offers a deployable option.

HAVELSAN's MAIN platform takes a different tack. Built around a model of roughly nine billion parameters, MAIN is an enterprise AI platform designed to run fully offline on closed networks—the kind of air-gapped deployment that matters when data cannot leave the building. HAVELSAN keeps adding modules (a recruitment module running on institutions' own servers was announced in October 2026). It is a preview of where regulated Turkish enterprises are heading: capable models that never touch a public cloud.

Related: Self-Hosted LLMs for Enterprises: When Local Models Beat the Frontier

The national AI action plan (2026–2030)

On 18 August 2026, the Türkiye AI Action Plan (2026–2030) took effect through Presidential Circular 2026/9, published in the Official Gazette (issue 33344). It replaces the 2021–2025 National AI Strategy and is coordinated by the Ministry of Industry and Technology. The plan has four axes, named in Turkish Fark Et, İstifade Et, Üret, and Yönet (notice, make use of, produce, and govern), with 16 actions between them.

The commitments most relevant to businesses:

  • AI vouchers: at least 1,000 AI vouchers for SMEs in the first 12 months.
  • Compute for everyone: a "GPU for Everyone" programme of compute credits for startups, SMEs, and researchers, starting at about 2 million GPU-hours a year and rising to 20 million a year by the end of 2028.
  • Data centres: at least 1 GW of AI and data-centre capacity by 2030, with a target of at least 10 billion dollars of investment, mostly private.
  • Regulatory sandboxes: at least five, in priority sectors such as finance, health, energy, mobility, and telecommunications.
  • A Turkic-languages LLM: a shared model covering the Oghuz, Kipchak, and Karluk language branches, built with Organization of Turkic States members, with a first version due by the end of 2027.
  • Governance: a National AI Board to monitor implementation, and a risk-based legal and ethical framework that is still to be drafted.

For a business, the practical reading is that compute, funding, and sector pilots for Turkish AI are becoming easier to access. The plan does not, by itself, create new obligations for private companies yet. Türkiye still has no AI-specific law: the AI bills submitted to parliament remain in committee.

Companies whose AI serves people in the EU already face binding rules, though: the EU AI Act applies to them even when they operate from Türkiye. See Does the EU AI Act Apply to Turkish Companies?

The commercial and enterprise models

Alongside the national effort, Türkiye's largest technology companies have adapted models to their own data and needs.

Turkcell released Turkcell-LLM-7b-v1 in 2024: Mistral 7B adapted with roughly five billion Turkish tokens using efficient fine-tuning methods (DoRA and LoRA), published under the Apache-2.0 licence. There has been no public successor model since, but Turkcell has shifted its weight to infrastructure: in July 2026 it set out a five-layer sovereign AI infrastructure strategy, backed by large data-centre investments.

Trendyol, the country's e-commerce heavyweight, has published a series of Turkish/English models on Hugging Face. They are fine-tunes of open models in the 7B to 12B range (built on Mistral, Llama 3, Qwen, and Gemma 3 bases), alongside a separate 70B cybersecurity model built on Llama 3.3. In 2026 it added open-weight models for its product catalogue, including a 27B vision-language model for catalogue quality and moderation, released under CC BY 4.0. For a retail platform, models tuned on product, review, and support language are a direct competitive lever.

VNGRS produced Kumru, one of the more talked-about efforts of the past year. Kumru was trained from scratch on Turkish, on roughly 500 GB of text. Kumru-2B was released with open weights under Apache-2.0 in autumn 2025, and the larger Kumru 7.4B launched at the same time as a demo and on-premises commercial product, without open weights. The training corpus itself has not been released. "From scratch" matters: rather than bolting Turkish onto an English base, VNGRS trained on Turkish first, which tends to produce more efficient tokenisation.

Earlier academic work such as Boğaziçi University's TURNA laid groundwork here, and the open-source Turkish NLP community—datasets, tokenisers, evaluation sets—remains the connective tissue that makes all of this possible.

A snapshot comparison

Model / project Who Status in October 2026 Best mental model for
BİLGE TÜBİTAK BİLGEM 1B and 9B from scratch, 27B continued pre-training, 122B in training; not yet generally available Public-sector alignment; one to watch
T3 AI T3 Foundation and Baykar Brief public beta in 2025; no public weights found A signal of sovereign intent
MAIN HAVELSAN ~9B model; runs fully offline on closed networks Air-gapped, regulated deployments
Turkcell-LLM-7b-v1 Turkcell Mistral 7B adapted to Turkish; open weights (Apache-2.0) A small, practical open baseline
Trendyol models Trendyol Turkish/English fine-tunes of open 7B–12B models; open e-commerce vision models E-commerce-tuned building blocks
Kumru VNGRS Trained from scratch; 2B open, 7.4B commercial Native Turkish tokenisation at small sizes

Read the last column, not the parameter counts. The right model is the one whose character matches your constraint—regulation, language quality, deployment environment, or budget—far more than the one with the biggest number.

What the benchmarks say about Turkish-specific models

A Turkish-specific model is not automatically better at Turkish tasks. The Cetvel benchmark, published in 2025, found that Turkish-centric instruction-tuned models generally underperformed multilingual and general-purpose models such as Llama 3 and Mistral. TurkBench, released in March 2026 by researchers from METU, Bilkent, Boğaziçi, Turkcell AI, and other institutions, covers 8,151 items across 21 subtasks—from knowledge and reasoning to Turkish grammar and instruction following—and runs a public leaderboard on Hugging Face.

The lesson is not that Turkish models are weak; small from-scratch models can be excellent at narrow tasks and are cheap to run on your own hardware. It is that "built for Turkish" is a hypothesis to test, not a guarantee. Use the public leaderboards to build a shortlist, then decide with your own documents.

What about the global chatbots?

None of this means the global players have left the room. In practice most Turkish enterprises still run production workloads on frontier international models, because raw capability, tooling, and reliability remain ahead at the top end. That is a legitimate choice—especially for internal, lower-risk automation where the data is not sensitive.

The shift is that it is no longer the only choice. Three forces are pushing Turkish organisations toward local or hybrid stacks: data residency and KVKK (sending personal data to a foreign API is a board-level question, not an engineering detail), cost at scale (once a workflow runs millions of calls, a right-sized local model can be dramatically cheaper), and language quality on genuinely Turkish tasks. Local infrastructure is also growing: Alibaba Cloud announced its first cloud region in Türkiye in September 2026, and new data-centre campuses are under construction. The mature answer for most is not "local or global" but a deliberate mix.

Related: GEO for Turkish Companies: How to Get Cited by ChatGPT and AI Search

What this means for your project

If you are choosing a foundation in 2026, a few practical principles hold regardless of which name is ascendant this quarter:

  • Start from the constraint, not the model. Regulated data pushes you toward closed or self-hosted deployment (the MAIN pattern). High-volume automation pushes you toward cost-efficient local models. Nuanced Turkish content calls for testing natively trained models such as Kumru alongside strong multilingual models. Let the hardest constraint choose.
  • Build only on what is available. Several national models are announced but not released. Plan around models you can download, license, or call today, and treat the rest as future options.
  • Design for model portability. The ecosystem is moving too fast to marry one model for good. Build your workflows—retrieval, tools, guardrails, evaluation—so the underlying model is a swappable component.
  • Evaluate on your own Turkish tasks. A small, honest test set of real questions with known-good answers will tell you more than any leaderboard.
  • Assume hybrid. Route each task to the model that fits it, and keep sensitive data on infrastructure you control while still reaching for frontier capability where it is safe to.
  • Use the incentives. AI vouchers, compute credits, and regulatory sandboxes under the 2026–2030 action plan can lower the cost of a first pilot. Check eligibility before you budget.

The takeaway

Türkiye in 2026 is no longer a market that simply consumes AI built elsewhere. It has state-backed model programmes, a from-scratch Turkish model family in Kumru, telecom and retail companies adapting open models, defence-grade closed-environment platforms, and a national action plan with measurable targets. Much of it, though, is still arriving: the flagship state models are not yet open, and the plan's compute and funding programmes are only starting.

For an organisation with a real project, that is good news with a caveat: more options and more leverage, but a harder design problem. Choosing well means understanding your constraints before your model.

At Orbitra, we help Turkish organisations design AI automation and agentic workflows that stay portable across this shifting model landscape—matching frontier, local, and self-hosted models to the right task, and keeping sensitive data on infrastructure you control. If you are mapping where to build, we can help you turn this landscape into a concrete architecture and a first pilot.

LET'S GET STARTED
PROJECT INTAKE

Let's put your AI automationinto orbit

Tell us the process, customer journey, or team capability you want to improve. We will map the right next step for AI automation consulting, an AI assistant, corporate training, or a measurable pilot.

Email Us[email protected]For inquiries, strategy, and demos
ORBITRA / PROJECT BRIEFSECURE CHANNEL
01
02
03
04