A reporter contacted me to comment on the recent launch of Turna.ai, as its answers reminded many of Grok’s scandalous answers. However, I could not follow the recent launch and thus could not respond to the reporter. But then I asked ChatGPT for a deep research. I believe users should not be harsh on this open-source initiative. Let’s see how it will go.

TURNA Turkish Language Model: TURNA is a state-of-the-art Turkish language model based on the UL2 encoder-decoder architecture, developed by the Computer Engineering Department at Boğaziçi University. It is designed to handle both natural language understanding and generation tasks in Turkish, outperforming many multilingual models and competing with strong monolingual models. TURNA was trained on a 10GB corpus from sources like OSCAR, OPUS, and Wikipedia, features 36 encoder and decoder layers, 16 attention heads, and over 1.1 billion parameters. It is distributed mainly for non-commercial academic research. Common research tasks for TURNA include text summarization, paraphrasing, news title generation, named entity recognition, semantic similarity, text classification, and more.arxiv+3

Introduction

Turna AI’s Kumru-2B model – a 2-billion-parameter Turkish large language model – was launched open source in October 2025, along with the extensive web corpus (500 GB of Turkish text) used to train it. This release (in both a base pre-trained version and an instruct fine-tuned version) quickly generated buzz in tech communities. Many observers hailed it as a milestone for Turkey’s AI ecosystem, noting it as “Türkiye’nin… ilk yerli LLM projesi” (the first locally-developed LLM) and a step toward digital independence[1][2]. On social media – especially Twitter (X) and Ekşi Sözlük – reactions ranged from enthusiastic praise and national pride to critical skepticism and humor. Below we break down the key sentiments, notable posts, comparisons, and industry commentary surrounding Kumru-2B’s debut.

Twitter Buzz and Positive Reception

On Twitter (X), many users and AI enthusiasts celebrated the open-source release. The model and its training corpus immediately drew attention – one AI commenter noted that Kumru-2B (both base and instruct) was released along with the web corpus it was trained on, a transparency move that impressed the community. This openness and focus on Turkish-language AI were widely praised. The model’s Hugging Face repository even trended on the front page shortly after launch, which Turkish tech observers took as a point of pride, signaling significant interest beyond Turkey. For example, one popular thread by a Turkish ML engineer highlighted “the value of open-sourcing a model” for the community and explained why Kumru-2B is an important development for Turkish NLP. Many congratulatory posts thanked the VNGRS/Turna AI team for contributing a Turkish-language LLM** to the open-source world, emphasizing how it advances local AI capability and sovereignty[2][3].

Notable Posts: Several influential figures amplified the news. The official VNGRS account shared Kumru’s release, and tech professionals like Merve Noyan and academic Akin Ünver retweeted it, calling the achievement “çok büyük gurur kaynağı” (“a great source of pride”) for Turkey. These posts received substantial engagement. Observers from abroad also took note; some English-language AI blogs described Kumru.ai as “a milestone in Turkey’s AI advancement”, underscoring its significance in demonstrating that smaller local teams can build useful LLMs from scratch[4]. Podcasts and tech forums discussed Kumru enthusiastically – Dijital Zihin, a Turkish AI podcast, praised the academic rigor of the project and its lead researcher Melikşah Türker for bridging local academia and industry[5]. Overall, on Twitter the sentiment was largely optimistic: pride in a home-grown model, appreciation for the open-source release (model and data), and hope that this would spur further AI innovation in Turkey.

Praise for Turkish Fluency and Use-Cases

Users who tested Kumru-2B highlighted its strengths in Turkish language tasks. Early adopters reported that Kumru’s outputs in Turkish felt fluent and natural, with nuanced understanding of Turkish grammar and idioms[6]. Some noted it outperformed larger multinational models on purely Turkish tasks, like summarization and grammatical error correction[6]. This resonated on social media, since many existing LLMs (e.g. GPT-4, Llama-2) are not Turkish-native and sometimes fumble with the language. Commenters lauded Kumru-2B for filling a niche: it can handle long Turkish documents (up to 8192 tokens context) and was even designed to run on modest hardware (fits on a 16GB GPU, or ~5GB for the 2B model)[7][8]. People mentioned use cases like document summarization and question-answering on Turkish texts as promising areas. The B2B focus was also understood – the team positioned Kumru not as a ChatGPT-style general chatbot, but as an on-premise solution for enterprises needing Turkish AI services (due to data privacy or regulatory reasons)[9][10]. Onedio, a popular Turkish media outlet, reported that Kumru “tanıtıldı, ilk günden büyük ilgi gördü” (“was introduced and received great interest from day one”)[11]. Indeed, many posts commended the VNGRS team for addressing a local need with a tailored model, and for planning future upgrades (a 7B model and multimodal abilities were hinted to be in the pipeline, fueling further excitement)[12].

Criticisms and Skepticism on Ekşi Sözlük

On Ekşi Sözlük, Turkey’s prominent online forum, the reaction was more mixed and often sarcastic. Some users were openly skeptical of the hype. The very first entry acknowledged Kumru as the first local LLM and “pretty fast”, but immediately noted “herhangi bir reasoning işlemi yok, internete bağlanamıyor – dümdüz LLM yapıyor” (“it has no reasoning, can’t connect to the internet – it’s just a plain LLM”)[1]. The post concluded that “şimdilik pek kullanılacak gibi değil”, i.e. it doesn’t seem very usable yet[1]. This set the tone for several other critical comments on Ekşi: users tested Kumru with basic questions and math and shared the amusing failures. One user quipped that Kumru is “inanılmaz yetenekli bir yapay zeka” in mathematics – because it confidently insisted that 3 + 3×5 = 9[13]. Another showed Kumru giving a nonsensical explanation of order of operations in math, humorously scolding the model: “yanlış özetliyorsun, kumrucum” (“you’re summarizing it incorrectly, my dear Kumru”)[14][15]. Such examples of hallucinations or mistakes spread on Ekşi, casting doubt on the model’s reliability.

Some commenters took issue with what they perceived as overblown claims. The “first national AI” branding was ridiculed by multiple users: “‘ilk yerli LLM projesi’ iddiasına götünüzle gülebilirsiniz”, one wrote, suggesting they could laugh their butt off at that claim[16]. They argued that many Turkish researchers and even students have been fine-tuning models, so Kumru isn’t the absolute first – just the first with heavy publicity. Another user pointed out they themselves had a custom LLM running locally with internet access and tools, so “akıllının biri çıkmış yapay zeka yaptım demiş” (“some smart aleck comes out and says ‘I made an AI’”)[17]. This cynicism highlights a sentiment that Kumru might be more of a rebranding or integration of existing open models rather than an earth-shaking innovation. In fact, one entry speculated that “LLM’ine sorabilirsiniz – Mistral ve LLaMA-3’ün merge edilmiş hali” (you can ask the LLM itself; it says it’s a merged Mistral+LLaMA-3 model)[18]. This may have been a misunderstanding – the developers later clarified that Kumru’s architecture is based on Mistral but trained from scratch, not a mere fine-tune or merge[19][20]. Still, the perception among some Ekşi Sözlük users was that Kumru didn’t feel entirely novel.

Despite the jokes and jabs, there were a few balanced voices on Ekşi Sözlük as well. Some recognized that Kumru is intended for enterprise use more than as a public chatbot, which explains its current limitations[21]. A user shared the VNGRS team’s official FAQ (posted in response to community questions) to address the criticisms: the FAQ explicitly notes Kumru has not yet undergone RLHF (reinforcement learning from human feedback) – analogous to OpenAI’s early GPT-3 (davinci-001) phase – and thus will make mistakes until further fine-tuning is done[22][23]. The team also answered why training from scratch (to get a Turkish tokenizer) was chosen over fine-tuning a foreign model, and emphasized that Kumru is not a ChatGPT rival but a smaller, domain-focused model[24]. After reading these clarifications, one Ekşi user conceded that at least “diğerleri gibi arkada ChatGPT API kullanıp ‘yerli milli yapay zekayız’ demiyorlar” – i.e. at least these folks are not secretly using the ChatGPT API in the backend while claiming to be a local AI[25]. This was a subtle nod to honesty; in the past, some so-called “Turkish AI” products were just wrappers around foreign models. In Kumru’s case, even skeptics acknowledged it is a genuine independent model – albeit one that “doesn’t know what it’s saying” yet[25]. In summary, Ekşi Sözlük’s reaction was a blend of patriotic enthusiasm dampened by technical realism – proud to see a Turkish model launched, but quick to mock its early shortcomings and any exaggerated marketing.

Comparisons to Other LLMs

Comparisons between Kumru-2B and other large language models were a recurring theme in discussions. Many drew parallels with OpenAI’s GPT series. Experienced users noted Kumru’s stage of development is roughly akin to GPT-3’s initial instruct-tuned model without alignment: the OpenAI davinci-001 (early 2022) level[22]. In other words, Kumru-2B can follow instructions but hasn’t been rigorously fine-tuned with human feedback yet, so it tends to hallucinate and err in predictable ways (e.g. struggling with math and factual precision)[26][23]. This was not meant entirely as a put-down – rather, it set expectations that Kumru will need further alignment to approach ChatGPT’s reliability. In fact, Kumru’s developers themselves made this comparison in their communications, saying “şu anda OpenAI’ın 2022 başında duyurduğu davinci-001 modeli ile aynı aşamadayız” (“we are at the same stage as OpenAI’s early-2022 davinci-001 model”)[27]. Knowing this, many users treated Kumru as a work in progress rather than a finished product, often explicitly stating that it’s not competing head-to-head with ChatGPT[10].

At the same time, the community did contrast Kumru with other available LLMs in terms of performance and purpose. The Cetvel benchmark results published by the team were frequently cited: remarkably, Kumru (both 7B and 2B) reportedly outscored much larger open models like LLaMA-70B, Qwen-72B, Gemma-27B, etc., on a suite of Turkish language tasks[28]. This claim garnered cautious admiration – if accurate, it highlights the benefit of a dedicated Turkish model. Some on Twitter found this result encouraging, pointing out that bigger isn’t always better if the data isn’t specialized; Kumru’s training on rich Turkish data gave it an edge in that niche[28]. However, others were quick to temper expectations by reminding everyone that ChatGPT (GPT-4) still vastly surpasses a 2B or 7B model in general capability. In fact, a number of commenters explicitly stated (or asked) if Kumru was a “ChatGPT’ye rakip” (a competitor to ChatGPT), to which the resounding answer – even from VNGRS – was “no”[24]. Kumru is smaller and cheaper by design, targeting on-premises deployment for Turkish companies, whereas ChatGPT is a massive, globally trained model hosted on the cloud[10]. Comparisons were also made with other local or regional LLM efforts: for instance, people mentioned models like Gemini or Grok (global models) and wondered how Kumru-7B (when released) might stack up, or referenced Korean and European “sovereign AI” projects in the context of countries developing their own models[4][29]. Overall, the consensus was that Kumru-2B is in the same league as other open small LLMs – useful within its scope, even outperforming larger models on Turkish tasks[6], but still far from the prowess of the top-tier models in broader domains[30].

User Concerns and Areas of Criticism

Beyond performance comparisons, social media users raised practical concerns about Kumru-2B’s current abilities. A major point of criticism was its tendency to “break character” or repeat boilerplate identity text. Developers testing the model noted Kumru would often respond with phrases like “Ben Kumru, VNGRS tarafından geliştirilmiş bir modelim…” (“I am Kumru, a model developed by VNGRS…”) in its answers, which became somewhat repetitive[31]. This was attributed to the instruct fine-tuning and lack of RLHF – essentially the model wasn’t deeply tuned to avoid self-identification in every reply. Similarly, reasoning ability was questioned. Apart from humorous math failures, some users found that Kumru struggled with understanding nuanced questions or implied context. One Twitter user gave a simple example: because they forgot a question mark, Kumru failed to realize a query was a question, whereas ChatGPT inferred it correctly – indicating Kumru’s conversational inference is still weak[32][33]. Hallucinations (making up facts) were also reported. Early users cautioned that Kumru might confidently provide incorrect information, a known issue in all LLMs but exacerbated here by the absence of a fine-tuning step to align outputs with factuality[26].

Math and logic were repeatedly pointed out as weak spots – an unsurprising outcome given that Kumru doesn’t have an internal calculator or reasoning module, and it hasn’t had techniques like chain-of-thought prompting or tool-use integrated. The developers openly acknowledged “Kumru matematikte kötü” (“Kumru is bad at math”) in their FAQ[34]. Social media users appreciated this honesty; several posts shared the explanation that, like early GPT models, Kumru simply predicts text and can’t inherently perform arithmetic or logical reasoning without training or tools[34][35]. Knowing this, some users tried to see if Kumru could be augmented – for instance, individuals talked about hooking Kumru up to Python tools or knowledge bases themselves, as a workaround. Indeed, one Ekşi Sözlük entry described how a hobbyist integrated a local LLM with Python functions and internet access, implicitly suggesting that Kumru could evolve in that direction too[17].

Another concern was the utility of the released web corpus. While many applauded that VNGRS open-sourced the 500GB Turkish web dataset used for training, a few developers on Twitter debated its quality and licensing. Could this corpus (presumably filtered and deduplicated Turkish web crawl data) become a valuable resource for others training models in Turkish? Enthusiasts saw it as a big win for the community, as it’s rare to have a large curated Turkish text dataset freely available. However, caution was noted: if much of that data is scraped from the internet, issues of data privacy or bias might need examination (though this topic was not front-and-center in the initial reactions). More commonly, users simply expressed curiosity: What’s in the corpus? Did it include social media posts, books, or just websites? Those kinds of discussions suggest that the open dataset release spurred interest among researchers who might repurpose or analyze it.

In summary, key concerns voiced by the public included Kumru-2B’s accuracy and maturity (needing RLHF to reduce hallucinations[31][36]), its reasoning limitations, and whether it’s currently more of a tech demo than a ready solution. Nonetheless, even critics often couched their comments with a sense of “it’s a start” – implying hope that these issues will be addressed in future iterations. The development team’s proactive approach (answering FAQs, promising upcoming improvements like a multimodal model and larger parameter versions) was noted and appreciated as a sign that feedback is being taken seriously[37][38].

Industry and Expert Commentary

The launch of Kumru-2B also sparked commentary from industry experts and AI researchers, both in Turkey and internationally. Industry insiders viewed Kumru through the lens of AI sovereignty and local ecosystem impact. A detailed analysis on one tech finance blog described Kumru.ai as a strategic leap for Turkey, emphasizing that having a domestically developed LLM (with open governance and Turkish data) is a step toward reducing reliance on US/China tech giants[4][3]. This narrative was echoed on LinkedIn and tech news, where people framed Kumru as part of Turkey’s effort to foster home-grown AI talent and infrastructure. The fact that VNGRS built this model for only around $250K using eight H200 GPUs over 45 days[39] was highlighted as an impressive feat – showing that relatively small teams can create meaningful models without billions of dollars. Experts pointed out how VNGRS even contributed code optimizations to open-source frameworks during this project (e.g. improvements merged into Hugging Face Transformers library)[40], underscoring the community-oriented approach.

Within Turkey, some AI practitioners expressed cautious optimism. Fatih Kadir Akın, a well-known Turkish developer, commented on X that Kumru’s current state is “kesinlikle… davinci-001’e eş değer” – essentially agreeing that it’s comparable to an unaligned GPT-3, which is understandable for now[41]. His remarks suggested that with further fine-tuning (and perhaps with community contributions), Kumru could become much stronger, just as GPT-3 evolved with InstructGPT and ChatGPT. Meanwhile, Deniz Oktar, the founder of VNGRS/Turna AI and leader of the Kumru project, took to social media to address the conversation. He thanked people for their interest and feedback, and in one post gently chided the harsher critics by noting that some of the criticisms revealed a lack of understanding about what it takes to develop and train an LLM[42]. Oktar emphasized that Kumru was a product of long-term research and engineering (not a weekend project, as some detractors implied) and invited experts to review it in depth. This kind of engagement from the creators – being transparent about the model’s development stages and receptive to feedback – earned respect from parts of the community. As one article put it, “VNGRS, Kumru’yla ilgili merak edilen sorulara cevap vererek ilgiye teşekkür etti” (“VNGRS answered questions about Kumru and thanked everyone for the interest”)[43].

Comparisons to global efforts: Some experts placed Kumru in context alongside other nations’ LLM projects. There was mention of how countries like Korea (e.g. SKT’s models) and organizations in Europe are also racing to build local-language LLMs[44]. In that frame, Kumru is seen as part of a broader trend of seeking “sovereign AI”. Turkish tech commentators noted that Kumru’s open-source release is aligned with European AI values (transparency, privacy, compliance with local laws like KVKK) and that it could serve as a testbed for responsible AI development in low-resource languages[45]. The academic community in Turkey also weighed in: researchers appreciated that VNGRS collaborated with university labs (e.g. Boğaziçi University’s NLP group) and built on prior local NLP projects like the TURNA encoder-decoder model[46]. This grounding in research lent credibility to Kumru. Several academics on Twitter and LinkedIn praised the release of the model weights and dataset as an invaluable resource for students and academics working on Turkish NLP, as they can now experiment on a state-of-the-art base instead of always relying on English-centric models.

In short, industry voices were largely supportive, framing Kumru-2B’s launch as an important experiment in building a Turkish AI ecosystem. The model is not yet a world-beater, but it’s a concrete starting point that can be iterated upon. The emphasis was on the long-term vision: if Turkey continues to invest in such projects (as VNGRS plans to with a 14B-parameter model by 2026[47]), it could yield not just better models but also local expertise and infrastructure for AI. As one analysis concluded, Kumru “may not yet rival GPT-class systems, but [its] nationally significant foundation positions it as a cornerstone in Turkey’s AI evolution.”[48]

Conclusion

The public reaction to Kumru-2B’s open-source release has been a vibrant mix of pride, excitement, critique, and humor. Turkish Twitter was abuzz with congratulations for achieving a first-of-its-kind local LLM and for sharing it openly, while forums like Ekşi Sözlük provided a reality check by poking at the model’s early flaws and hype. Key themes emerged:

Gratitude and National Pride: Many see Kumru as a meaningful step toward technological self-reliance, applauding the open-source ethos and the focus on Turkish language capabilities[4][3].
Constructive Criticism: Users did not hold back in pointing out Kumru-2B’s limitations – from factual inaccuracies and math errors to repetitive responses – often with wit and memes[49][16]. This critical feedback is coupled with advice (calls for RLHF, better reasoning) and patience for future improvements.
Comparisons and Context: The model’s launch prompted discussions about where it stands relative to giants like ChatGPT and other open models. The consensus is that Kumru is not a ChatGPT killer, but it fills a gap for Turkish NLP, outperforming some larger models in that niche while clearly trailing top-tier AI in general ability[30][10].
Community and Ecosystem Impact: Perhaps most importantly, Kumru-2B’s release has galvanized the Turkish AI community. By open-sourcing not just the model but also the training data, Turna AI/VNGRS invited developers, researchers, and even skeptics to tinker with and improve the model. This collaborative spirit has been widely recognized as the true victory – seeding knowledge and tools that can spur further innovation. As one commentator noted, Kumru’s story “illustrates what disciplined innovation and long-term vision can achieve”[50], even if the first step is small.

Going forward, the conversations suggest that Kumru’s success will be measured by how it evolves. The initial social media buzz has died down somewhat, but the project has entered the phase of incremental improvement with community feedback. If the team addresses the critiques (e.g. by implementing RLHF to curb hallucinations, and releasing the promised larger models), public sentiment may tilt even more positively. For now, Kumru-2B’s launch stands as a landmark moment that combined local tech ambition with open-source collaboration, eliciting reactions that are as insightful as they are passionate – truly reflective of a community invested in its AI future.

Sources: The above analysis is based on a range of public reactions and commentary, including Turkish social media discussions on Ekşi Sözlük[1][49][16], Twitter posts and threads from AI enthusiasts and the Kumru team[31], as well as expert commentary from blogs and news outlets[4][6]. These sources provide a multi-faceted view of how Kumru-2B’s open-source debut was received both within Turkey and in the broader AI community.

[1] [9] [10] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [24] [25] [26] [27] [32] [33] [34] [35] [37] [38] [49] kumru.ai – ekşi sözlük

https://eksisozluk.com/kumru-ai–8034656

[2] [3] [4] [5] [6] [30] [31] [36] [39] [40] [45] [46] [47] [48] [50] Kumru.ai: Turkey’s Strategic Leap in AI Sovereignty — Verified Insights Beyond the Headlines | Kurumsal Finans ve Strateji Rehberi | Finance & Strategy Insights