Researchers propose Mage-VL-Cowork, a 4-billion-parameter hybrid computer-use agent trained via a five-stage curriculum
The agent combines visual GUI interaction, code execution, and terminal commands, achieving 66.0% on OSWorld-Verified while reducing per-task training costs by 82.6%.
1 source · cross-referenced
- Mage-VL-Cowork is a 4B-parameter hybrid computer-use agent combining visual GUI interaction, code, and terminal execution.
- Training uses a five-stage progressive curriculum across midtraining, SFT, and RL to reach 66.0% on OSWorld-Verified.
- Mage-CUA-Forge synthesizes 293,806 desktop and web tasks across 6,915 skills, cutting recurring per-task costs by 82.6%.
- The approach evolves teacher harnesses and instantiates specialized CUA harnesses to generate high-quality trajectories.
Researchers from multiple institutions introduce Mage-VL-Cowork, a 4-billion-parameter hybrid computer-use agent designed to combine visual GUI interaction with code execution and terminal commands within a single policy. The model is trained from a general multimodally aligned checkpoint using an end-to-end, progressive five-stage curriculum spanning midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL).
To scale the curriculum, the team developed Mage-CUA-Forge, a framework that replaces per-task agent construction with reusable, skill-conditioned generation workflows and application-specific preparers. This system synthesizes 293,806 desktop and web tasks across 6,915 skills and reduces recurring per-task expenditure by 82.6%.
Rather than distilling from a single default teacher, the authors evolve the teacher harness and instantiate specialized computer-use agent harnesses to produce high-quality trajectories alongside stage-specific supervision. The resulting model, Mage-VL-Cowork-4B, achieves 66.0% on OSWorld-Verified after progressing through the training stages, reporting more efficient reasoning costs and state-of-the-art performance at its parameter size.
The work targets on-device intelligence, emphasizing efficiency and scalability in training computer-use agents that can generalize across tasks without bespoke engineering for each scenario.
- Nov 11, 2026 · Latent Space — swyx
OpenAI demonstrates voice steering, cloud handoffs, and personal agents for Codex at AI Engineer Code Summit
Trust79 - Nov 11, 2026 · The Verge — AI
OpenAI to launch verified AI Business Agents in ChatGPT for US e-commerce and service businesses
Trust79 - Nov 10, 2026 · Latent Space — swyx
NaiveAI releases Naive-N0.5-Flash coding model with ultrafast token throughput
Trust74