A small Tajik model increases its tool call score from 20.69 to 64.14
The model works on small computers. Its score on a general English knowledge test went down from 60.8 to 52.5.
The authors released Saidzoda-Tajik-LFM-2.6B, a Tajik version of LFM2.5-2.6B-Base, a model with 2.6 billion parameters. It makes tool calls, which are messages from a model to a tool, in Tajik on small computers. On a test of 174 items, the strict score on Tajik prompts went from 20.69 to 64.14. The first model had 20.69.
How it works
The work had two steps. The first step was more pretraining on 1,485,319 cleaned Tajik web documents. The second step used machine-translated Tajik tool calls, English tool calls and Tajik instructions.
The numbers
The increase is 43.45 points. English tool call scores did not change. The score on a Tajik knowledge test went from 28.0 to 35.0. The score on English MMLU, a general knowledge test, went from 60.8 to 52.5.
What is not known
The model is not a general Tajik assistant. It is right on only 6.9% of the cases when no tool fits. The authors tell that a person must check each call before the system makes it.
Insight: A repair for a small language has a price
On 26 October 2026, BabelArena found that low-resource languages gave 5 frontier models more tool errors. This paper closes a part of that difference for Tajik. But the checkpoints after the first step had a score of zero on tool calls. The English test went down. A small team can close a language difference, and local builders look like they have the control. Each repair costs some general skill.
Sources
Posted
