Tajik, the state language of Tajikistan, is written in Cyrillic with six letters that Russian does not have, and open models almost never list it among their supported languages. Last Friday, Saidzoda AI Lab in Dushanbe posted Saidzoda-Tajik-LFM-2.6B to arXiv: a 2.6B model built specifically for tool calling in Tajik, with a 12-page technical report attached. Call centres, government services and banking never wanted chat to begin with — they need a small model that turns a Tajik request into one correct call to an existing system. That is exactly the job.
The recipe: two LoRA stages, 3 H100s for 47.1 hours
The base is Liquid AI's LFM2.5-2.6B-Base, a hybrid of short convolutions and grouped-query attention. Stage one is LoRA continued pretraining (CPT): 1,485,319 cleaned Tajik web documents for 3 epochs and 17,574 optimiser steps, 96 packed sequences of 4,096 tokens each, on 3 H100 GPUs for 47.1 hours; training loss fell from 6.615 to 1.662. Stage two is SFT: machine-translated Tajik tool calls, 61,798 untranslated English tool calls kept as replay, and Tajik instructions, mixed at 40/35/25 for one epoch only, trained with axolotl throughout.
The result: Tajik 20.69 to 64.14, English untouched
The test set is their own 174-item Tajik translation of BFCL v3 (89 single, 27 parallel, 29 irrelevance, 29 multi-turn). Strict full-call accuracy: the base answers 76.55% of English items, but only 20.69% when the same request is written in Tajik. The adapted model lifts Tajik to 64.14, a paired gain of +43.45 points (95% interval +35.2 to +51.7), while English sits at 77.24. A lenient parser that also accepts the base's own call formats scores the base at 31.72 — the gap is real, but it narrows considerably under the lenient yardstick.
The cost: CPT zeroes tool calling to 0.00, and the model almost never says no
The most informative finding sits in the stage-by-stage measurement: every CPT checkpoint scores 0.00 in both languages. Continued pretraining on raw web text wipes out the base model's native tool-call format entirely — and, as the lenient re-scoring shows, part of the skill itself. SFT with English replay rebuilds the format and the English skill together, but the bill is long: English MMLU falls from 60.8 to 52.5; when no offered tool fits, correct refusal collapses from 89.7 to 6.9 in Tajik and from 62.1 to 31.0 in English — the model now reaches for a tool on almost every request; multi-turn calls stay weak (3.45 to 17.24). The model card says it plainly: its calls should be checked before they are executed. On the plus side, the Tajik knowledge benchmark rises from 28.0 to 35.0, while Belebele reading comprehension moves 37.6 to 40.0 with no significant change.
So what
The paper states, in its own words, that to the authors' knowledge this is the first public Tajik tool-calling model released with a technical report — the authors searched arXiv and the Hugging Face Hub in October 2026 and found none. For long-tail language communities the template is worth more than the model itself: data sources, configurations, compute, training logs and every evaluation are published, and the eval_results folder ships the original files. The weights are merged bfloat16 at about 5 GB under the LFM Open License v1.0 — commercial use is capped by a revenue threshold, and downloading requires filling a form at saidzoda.com first. Training compute came through the NVIDIA Inception programme, and the acknowledgements note that Anthropic Claude helped write scripts and draft text while the authors checked every number. The broader lesson: injecting a new language into a small model is not "translate data and fine-tune" — CPT zeroes the instruction and tool format first, and English replay is the guardrail that brings it back. The other half of the lesson is unfinished: teaching a model to call tools also taught it to always call. If a low-resource language wants agents, someone has to watch that 6.9%.
References: arXiv:2610.41576 · HF model card