Event date · · Qwen (Alibaba)

Qwen releases RecreationWorld: scalable verifiable environments for hybrid computer-use agents

Alibaba 阿里巴巴Chinese AIOpen weights
FACT STATEMENT

Qwen (Alibaba) open-sourced RecreationWorld on GitHub under the MIT License. It is a five-platform framework for hybrid computer-use agents that interleave GUI exploration, coding, and visual verification. RecreationBench provides 250 held-out tasks with programmatic and visual evaluation. Reported benchmark scores: GPT-6 Astra 58.06 average, Claude Opus 5 44.16, GPT-5.6 Sol 42.06, Grok 4.6 36.73, Qwen3.8-Max-0902 34.80, Kimi K3 31.41, Claude Opus 4.8 31.10, GLM-5.3 24.38, Gemini 3.7 Flash 21.12, Qwen3.7-Plus 9.15. Estimated costs per task range from $1.23 (Qwen3.7-Plus) to $117.17 (Claude Opus 5).

China context

Original name
通义千问
Outside China
Open weights · github.com
Claims
Company-reported; not yet independently evaluated
For builders
Builders outside China can use RecreationWorld to train and evaluate computer-use agents on five platforms with a permissive MIT license.
For investors
Investors can track adoption of RecreationBench as a standard for agent evaluation and monitor Qwen's progress in agentic AI.
What happened

Qwen (Alibaba) released RecreationWorld, an open-source framework for training and evaluating hybrid computer-use agents that can explore GUIs, write code, and visually verify their work. It includes RecreationBench, a benchmark of 250 tasks across Ubuntu, macOS, Windows, Android, and Web platforms. The repository provides setup guides and evaluation interfaces, and is licensed under MIT.

Technical significance

RecreationWorld frames tasks as recreating a running reference application, using the reference as an executable oracle. Agents loop through explore-implement-verify, and scoring is based on observable behavior via programmatic and visual assertions, not source similarity. This allows cross-language and cross-framework implementations.

Industry impact

The benchmark results show a wide performance gap between frontier models, with GPT-6 Astra leading at 58.06 average and Qwen's own Qwen3.7-Plus at 9.15. Cost per task varies dramatically, from $1.23 to $117.17, highlighting trade-offs between capability and cost for computer-use agents.

Decision value

For developers, RecreationWorld provides a free, open-source environment to train and test computer-use agents across multiple OS platforms. For enterprises, it offers a way to benchmark agent capabilities and costs before deployment. The MIT license allows commercial use.

What to watch

Observable next signals include whether other labs adopt RecreationBench for agent evaluation, whether Qwen releases trained agent models based on RecreationWorld, and whether the benchmark is integrated into leaderboards or agent frameworks.

CHINA AI WEEKLY

Get the week in Chinese AI, in English.

One weekly issue of verified model, company, robotics and policy changes, each with its original source and outside-China availability.