# Tencent and Fudan University Launch WorkForge Framework for Agent Training Environment Construction

- Category: Flash (快讯)
- Published: 2026-10-23T11:56:00.000Z
- Source date: 2026-10-23
- HTML: https://www.88z.ai/en/news/flash-20261023-63253f467fa0f83b
- Markdown: https://www.88z.ai/en/news/flash-20261023-63253f467fa0f83b.md
- Other language: https://www.88z.ai/news/flash-20261023-63253f467fa0f83b.md

Tencent Hunyuan and Fudan University jointly launched the WorkForge framework, designed to scale the construction of verifiable agent work environments from real-world resources. They simultaneously released the WorkForge-16K dataset, covering 16,700 work environments across 40 professional fields. The WorkForge-27B model, trained on real workspace trajectories from this framework, performed excellently, comprehensively surpassing models with hundreds of times more parameters in five major work

## At a glance

- Tencent and Fudan launched WorkForge framework and WorkForge-16K dataset with 16,700 environments
- WorkForge-27B model outperforms models with 100x more parameters on five major benchmarks

## The update

Tencent Hunyuan and Fudan University collaborated to launch the WorkForge framework, which is dedicated to scaling the construction of verifiable agent work environments using real-world resources. Accompanying the framework is the WorkForge-16K dataset, containing sixteen thousand seven hundred work environments spanning forty professional fields. Building on this, the team trained the WorkForge-27B model using real workspace trajectories. Test results show that the model performed excellently on five major work agent benchmarks, including GDPval, comprehensively surpassing models with a hundred times more parameters, demonstrating the framework's effectiveness in enhancing agent work capabilities.

## Sources

- [腾讯混元](https://mp.weixin.qq.com/s/GYyQ5fS4HMBK8qoNcPJlng) (2026-10-23)

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.