Illustration: AI
Kazakhstan already has at least four large language models that understand Kazakh: KazLLM, AlemLLM, the Oylan assistant and Sherkala, which was built in the UAE. According to the Ministry of Digital Development, the two main models, KazLLM and AlemLLM, power 42 AI products. Yet in February, the president asked the minister why people still turn to ChatGPT. Here’s what these models are, how they differ and where they’re already being used.
KazLLM: the first model, with 8 billion and 70 billion parameters
KazLLM was built by Nazarbayev University’s Institute of Smart Systems and Artificial Intelligence (ISSAI). Development ran from April to December 2024, and the model was presented to the president on December 11, 2024 (issai.nu.edu.kz, tengrinews.kz).
The model wasn’t built from scratch. ISSAI took Meta’s open Llama 3.1 architecture and trained it further on Kazakh data. It released two versions: an 8-billion-parameter model for workstations and a 70-billion-parameter model for data centres. Training used eight NVIDIA DGX H100 servers supplied by a private telecoms company. According to ISSAI, at launch, the 70-billion-parameter version came close to matching GPT-4o on aggregate benchmark scores (issai.nu.edu.kz).
According to Tengrinews, more than 140 specialists from 26 institutes and universities took part, and the corpus covers 115 disciplines. The model understands Kazakh, Russian and English.
Beeline Kazakhstan and QazCode describe themselves as KazLLM’s developers “in partnership with ISSAI and Astana Hub” and report a corpus of more than 150 billion tokens in four languages, including Turkish (standard.kz). In October 2026, the project won an AmCham Award in the “Kazakh Culture” category.
The model is publicly available on Hugging Face, but under a non-commercial licence: you can use it for research, but you can’t sell a product built on it. In late 2024, the university granted Astana Hub a non-exclusive licence. ISSAI says it has not been asked to continue development since then.
AlemLLM: the largest, with 247 billion parameters
The second model, AlemLLM, arrived in summer 2025. Astana Hub released it with technical support from 01.AI, the Chinese company founded by Kai-Fu Lee (huggingface.co).
This model operates on a different scale: 247 billion parameters, with 22 billion active for each response, using a “mixture of experts” architecture. Its Apache 2.0 licence allows use in commercial products.
In the developer’s own tests, AlemLLM scores an average of 0.826 in Kazakh against GPT-4o’s 0.776, and 0.848 in Russian against 0.808. There’s an important caveat: these are Astana Hub’s own measurements, and GPT-4o has long since ceased to be OpenAI’s strongest model. GPT-4o still leads in some individual tests, including HellaSwag in all three languages (huggingface.co).
Oylan: an assistant you can try
While KazLLM and AlemLLM are “engines” for developers, Oylan is a ready-to-use assistant for everyday users. ISSAI is developing it too. Its third version was unveiled in December 2025: it handles text, voice and images, understands Kazakh (including Latin script), Russian and English, and can search the web (issai.nu.edu.kz).
You can try it at oylan.nu.edu.kz. ISSAI also has a whole range of tools: MangiSoz recognises and synthesises speech in 11 languages, TilSync translates speeches in real time, and Beynele generates images after training on Kazakh visual data.
Sherkala: a Kazakh model from the UAE
Sherkala was built outside Kazakhstan. It was released in February 2025 by Inception, part of the UAE’s G42 group, together with MBZUAI university and Cerebras Systems (huggingface.co).
It’s a small model with 8 billion parameters, also based on Llama 3.1. It received further training on 45 billion tokens, and its vocabulary was expanded by 25% with Kazakh words. On the developers’ Kazakh benchmarks, Sherkala scored 47.6 against 43.7 for the same-sized KazLLM. Its licence prohibits commercial use.
42 products and 480 thousand requests
Where is all this being used? According to the Ministry of Digital Development, as of February 2026, 42 AI products had been built on KazLLM and AlemLLM, and the models had received more than 480 thousand requests: 420 thousand from private businesses and 60 thousand from the public sector. The model had been downloaded from Hugging Face by 135 companies in 13 countries (kz.kursiv.media).
Examples include the Qazaq law legal adviser, a consular agent for Kazakhstanis abroad, procurement and tax agents, and AI eGov solutions for specific sectors. Samruk-Kazyna has introduced three AI agents, including SK AI, a “digital equivalent of a board member”. According to Beeline, KazLLM also powers commercial services: AI Tutor, Mentra, Aventa and the AIDA medical assistant (standard.kz).
Why people still turn to ChatGPT
At a government meeting on February 11, 2026, Tokayev asked Minister of Digital Development Zhaslan Madiyev about weak demand for domestic models: users prefer foreign services. The minister replied that the models had been integrated into more than 40 information systems and were available to students, startups and IT companies (tengrinews.kz).
Experts point to several reasons. There are no public comparative benchmarks or regular progress reports, and many Kazakhstanis simply don’t know about the projects (tengrinews.kz). ISSAI itself points to the resource gap: Meta had more than 16 thousand servers of the same type and 400 researchers working on Llama, while Kazakhstan spends around 0.2% of GDP on science, compared with 1% or more in developed countries (issai.nu.edu.kz). New models appear globally roughly every six months. Without continued training, a national model quickly falls behind.
There’s a simple explanation too: global services have also learned Kazakh. Google announced Kazakh language support in Gemini, while OpenAI and Anthropic models answer in Kazakh without any special setup.
What’s next: 14 billion tokens with OpenAI
The next step is data. On November 9, 2026, Rauan Kenzhekhanuly, chairman of the Qazaq Tili society, told the president that the society and OpenAI had collected a text corpus of 14 billion tokens and 12 thousand hours of audio recordings in Kazakh (tengrinews.kz). Both global and local models need these corpora: the more high-quality Kazakh text there is, the better any AI model understands the language.
The Ministry of Digital Development has also outlined other plans: a voice model and small models for specific industries that can work offline, for example in mines (kz.kursiv.media).
Have you tried Kazakh AI models? Which do you use: Oylan, ChatGPT or something else?
