# Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model

Source: [arXiv](https://arxiv.org/abs/2610.32933v1)  
Feed7 permalink: https://feed7.dev/p/2610-32933v1-0kbcu6k  
Published: 2026-10-23T17:30:40.000Z  
Trust: Needs Review (needs_review)

## Why Included

Yalda Embedding offers a compact Persian retrieval model with released weights, data and code. Its results suggest distillation can improve retrieval without increasing inference size.

## Source Summary

Yalda Embedding adapts a **595.8M-parameter** model with over **1 million query-document pairs**. On eight FaMTEB tasks, its macro-average is **0.5488**, the highest among evaluated 0.6B models.

## Practical Implication

Builders adding Persian retrieval can evaluate the released model and inspect the dataset and training code. The final distillation step raised the mean from 0.4813 to 0.5488 without increasing inference-time model size.

## Agent-Ready Context

Yalda Embedding adapts a **595.8M-parameter** model with over **1 million query-document pairs**. On eight FaMTEB tasks, its macro-average is **0.5488**, the highest among evaluated 0.6B models.

Builders adding Persian retrieval can evaluate the released model and inspect the dataset and training code. The final distillation step raised the mean from 0.4813 to 0.5488 without increasing inference-time model size.

Results are specific to the reported FaMTEB tasks and one-epoch setup. The paper reports comparisons against selected models, not broader production performance.

## Connected Context

Feed7 judgment across 910 accumulated Signals:

This adds Persian retrieval evidence for a compact open model, with the reported distillation gain improving benchmark performance without increasing inference-time size. It makes language-specific retrieval a concrete model option to evaluate, while keeping the case bounded to the reported FaMTEB tasks and setup.

- [superlinked/sie](https://feed7.dev/p/sie-055bka3) — Yalda supplies a specialized retrieval model that could fit the kinds of task-specific models SIE serves behind a unified API; the paper does not evaluate integration or serving through SIE.
- [When Does Bigger Help? A Controlled Study of LLM Scale for Ontology Learning](https://feed7.dev/p/2608-31118v1-1abyqwd) — Both report that model choice should be judged on the target task: Yalda’s result is specific to Persian retrieval benchmarks, while the ontology study finds size rankings vary by task.

## Context Map

- Layer: model
- Domains: research, data
- Topics: retrieval, open-models

## Uncertainty

- Results are specific to the reported FaMTEB tasks and one-epoch setup. The paper reports comparisons against selected models, not broader production performance.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.