RedOne 2.0: Xiaohongshu Next-Generation SNS Scene Large Model Analysis

DreamActor Team 2025-11-18 6 min read

RedOne 2.0: Xiaohongshu Next-Generation SNS Scene Large Model Analysis

Xiaohongshu's new self-developed model RedOne 2.0 has appeared on arXiv, quickly sparking discussions in the tech community. Although the official announcement has not been fully released, the paper and public evidence confirm that RedOne 2.0 is a major upgrade specifically designed for social network scenarios (SNS), covering typical tasks such as content understanding, recommendation, moderation, and dialogue.

This article summarizes the characteristics, technical architecture, and potential applications of RedOne 2.0 based on currently available information.

What is RedOne?

RedOne is Xiaohongshu's self-developed language model dedicated to social network tasks, focusing on solving common text, semantic, and scenario problems on the platform, including:

  • Content understanding (comments, posts, titles)

  • Tagging and intent classification

  • Multilingual dialogue

  • Search query understanding

  • Risk content detection

    and more.

The previous generation model adopted a structure of "continued pre-training → supervised fine-tuning → preference optimization (RHO)", with customized training for typical SNS scenarios.

RedOne 2.0 Comprehensive Upgrade Highlights

1. Two Model Versions (Exposed by Community)

According to the paper and community technical leaks, two version information has emerged:

RedOne 2.0-4B

  • Lightweight version

  • Focuses on low-cost inference and deployment

  • Known tests show: 4B-level models exceed common 7B baselines in some tasks

RedOne 2.0-30B-A3B

  • High-performance version

  • Suitable for large-scale text understanding and moderation tasks

  • "A3B" may represent an upgraded three-stage training system

Note: The above version numbers come from the paper and technical community, and have not been officially announced.

2. New ETR Training Paradigm

RedOne 2.0 introduces the ETR (Exploratory → Targeted → Refinement) training process, making the model more suitable for "high-noise, fragmented, multi-scenario" SNS data.

(1) Exploratory Learning

  • Uses large-scale unstructured social data

  • Does not rely on labels

  • Model learns real user expression methods, tone, and style

(2) Targeted Fine-Tuning

  • Covers tasks such as comment classification, content recognition, and query parsing

  • Precisely aligns with actual business scenarios

(3) Refinement Learning

  • Similar to RLHF

  • Emphasizes "social text preferences" rather than general dialogue preferences

  • Makes generated content closer to user communication styles

This system makes RedOne 2.0 perform more stably and accurately in specific application areas.

3. Performance Metrics

According to evaluations disclosed in the paper:

  • 4B model exceeds mainstream 7B-level models by about 2.41 points in some key tasks

  • Significant improvements in content understanding, tag prediction, multilingual intent parsing, and other tasks

  • Stronger understanding of social context, emotions, and colloquial expressions

Although current public data is limited, the model's task design already demonstrates advantages in the SNS customization direction.

Potential Landing Directions for RedOne 2.0

Combining the model structure and task adaptation approach, RedOne 2.0 may serve the following areas:

1. Content Moderation

  • Text risk detection

  • Emotion and context recognition

  • Scenario-based violation judgment

  • Moderation assistance explanation

2. Search Enhancement

  • Search intent understanding

  • Query rewrite

  • Multilingual retrieval

  • Semantic embedding optimization

3. Content Creation Assistance

  • Title generation

  • Copy optimization

  • Image-text note content suggestions

4. Recommendation System Enhancement

  • Topic extraction

  • Interest tag generation

  • High-dimensional semantic embedding generation

5. AI Interaction Features

  • Intelligent comment assistant

  • AI store dialogue assistant

  • Automatic user question replies

These application scenarios are highly consistent with RedOne's training direction.

Summary

RedOne 2.0 is a large model upgrade version focused on SNS scenarios, focusing on solving real complex tasks in the social content field.

  • Introduces ETR three-stage training system

  • Significantly enhanced performance compared to the previous generation

  • 4B model exceeds common 7B models in multiple tasks

  • Has stronger social context understanding capabilities

  • May be deeply integrated into Xiaohongshu's content moderation, search, recommendation, and interaction systems

As the paper is published and more details are revealed, RedOne 2.0's specific capabilities, open methods, and actual productization will become increasingly clear.

Back to Blog