# Xiaohongshu Open-Sources BigMac, a Pipeline-Parallel Training Framework for Multimodal Large Models

- Category: Flash (快讯)
- Published: 2026-07-24T11:54:00.000Z
- Source date: 2026-07-24
- HTML: https://www.88z.ai/en/news/flash-20260724-f70b15a65b7b0250
- Markdown: https://www.88z.ai/en/news/flash-20260724-f70b15a65b7b0250.md
- Other language: https://www.88z.ai/news/flash-20260724-f70b15a65b7b0250.md

Xiaohongshu's dots infra team has open-sourced BigMac, a multimodal large model training framework that balances computational efficiency and memory usage via nested pipeline design, achieving significant speedups and already deployed in production.

## At a glance

- BigMac uses a nested pipeline design with an LLM backbone to embed encoding and generation, balancing efficiency and memory.
- Training speed improves by 1.08 to 1.9 times over baselines with stable memory usage, now used in dots series production.

## The update

The dots infra team at Xiaohongshu has officially open-sourced BigMac, a training framework designed for multimodal large models. This framework addresses the technical challenge of balancing computational efficiency with memory usage by introducing a dependency-safe nested pipeline design. BigMac uses the large language model pipeline as its backbone, orderly embedding encoder and generator computations to optimize overall resource scheduling. Experimental results show that BigMac increases training speed by 1.08 to 1.9 times compared to baseline methods, while maintaining stable memory consumption. Currently, this efficient training framework has been successfully applied to the production training of Xiaohongshu's dots series models, demonstrating its effectiveness and reliability in real-world business scenarios.

## Sources

- [小红书技术REDtech](https://mp.weixin.qq.com/s/tNJARtn1jIayL5URg87Row) (2026-07-24)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
