# Yalda researchers release Persian retrieval model and training data

_Published Monday, October 26, 2026 at 1:28 PM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers released Yalda Embedding, a 595.8-million-parameter model for finding Persian documents relevant to a query, alongside model weights, training and evaluation code, and datasets. Their preprint reports a 0.5488 macro-average across eight Persian retrieval tasks, outperforming the evaluated 0.6B models on every task.

The model adapts Qwen3-Embedding-0.6B using more than one million query-document pairs. Training with guidance from a larger model raised its mean score from 0.4813 to 0.5488 without increasing inference-time model size. The released data include those teacher-generated representations, allowing reuse without repeating the teacher inference. Results remain limited to the reported evaluations.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.32933)

---
Canonical: https://techandbusiness.org/newswire/ANgHwzOZZVEBetJepzFILz
Published: 2026-10-26T17:28:00.949Z
Story chronology: 2026-10-26T04:00:00.000Z
Retrieved: 2026-10-26T20:02:13.771Z
Publisher: Tech & Business (techandbusiness.org)