Sherkala-Chat
Open LLM for Kazakh language tasks
Sherkala-Chat is an instruction-tuned large language model designed for Kazakh speakers. Built on LLaMA-3.1-8B by a research team, it was released in 2025 to improve inclusivity for moderately resourced languages.
- Built on LLaMA-3.1-8B architecture
- Trained on 45.3 billion tokens
- Optimised for Kazakh language
- Released with open weights
- Includes manually verified local data
What it is
Sherkala-Chat is a generative large language model focused on the Kazakh language. It contains eight billion parameters and operates as an instruction-tuned assistant. The model aims to support Kazakh speakers by providing strong reasoning and knowledge capabilities in their native tongue. It also maintains competitive performance in English while prioritising Kazakh inclusivity.
How it came about
The model was developed by a large group of researchers including Fajri Koto and Preslav Nakov. They adapted the existing LLaMA-3.1-8B architecture to better serve Kazakh users. The team published their work in March 2025 and revised it in October 2025. Their goal was to address the lack of high-quality models for moderately resourced languages.
Inside the model
Training involved 45.3 billion tokens drawn from Kazakh, English, Russian and Turkish sources. The developers used translated instruction datasets alongside a specific dataset for Kazakhstan. This local data was automatically constructed and then manually verified. Safety alignment relied on Kazakh-specific data to ensure responsible outputs. The base structure remains a decoder-only transformer derived from LLaMA.
What it can do
It demonstrates strong knowledge and reasoning skills in Kazakh. Benchmarks show it significantly outperforms other open Kazakh and multilingual models of similar size. In English tasks, it achieves competitive results compared to peers. The model is designed to handle general instruction following while maintaining linguistic accuracy for Kazakh queries.
How to use it
The weights are released openly for public use. This allows researchers and developers to build real-world applications for Kazakh speakers. The team provides detailed documentation on training methods and evaluation metrics. No proprietary restrictions limit access to the model parameters.
Safety and testing
Alignment processes included Kazakh-specific safety data. The team leveraged both translated datasets and locally verified instructions. This approach aimed to ensure the model behaves responsibly within the cultural and linguistic context of Kazakhstan.
Why it matters
The release supports research into low-resource language modelling. It provides a baseline for future Kazakh natural language processing tasks. By sharing training details, the authors encourage further development in regional AI inclusivity.
Questions people ask about Sherkala-Chat
What languages does Sherkala-Chat support?
The model is primarily designed for Kazakh but was trained on data in Kazakh, English, Russian and Turkish. It performs strongly in Kazakh and competitively in English.
Is Sherkala-Chat free to use?
Yes, it is released as an open-weight model. This means the parameters are available for download and use in research or commercial applications without proprietary barriers.
How large is the Sherkala model?
It has eight billion parameters. This size balances computational efficiency with strong performance in reasoning and knowledge tasks for its target languages.
Good reading and sources
There is no English Wikipedia article on Sherkala-Chat yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2503.01493. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.