Alibaba scales Qwen agent to 397B parameters for native software control
Alibaba's Qwen team has built a 397-billion-parameter mixture-of-experts agent capable of operating arbitrary software through raw screenshots and input events, without relying on structured APIs or accessibility layers. The system maintains visual context across up to 20 frames and compresses older history into fixed blocks, enabling long-horizon task execution. Training leveraged a massive cloud rollout infrastructure spanning nearly 100,000 vCPUs and concurrent environments. This represents a significant step toward general-purpose computer automation, moving beyond domain-specific agents and raising the bar for what foundation models can accomplish in unstructured, real-world software environments.
Modelwire context
Analyst takeQwen-CUA operates on raw pixels and input events rather than APIs, which sounds like generality but actually sidesteps the harder problem: how agents verify their actions succeeded. The 20-frame context window and history compression are engineering choices that trade off long-horizon coherence for computational tractability, not breakthroughs in reasoning.
This launch arrives the same day as Qwen3.8-Max, signaling Alibaba is pursuing both scale (2.4 trillion parameters for reasoning) and breadth (native computer use across arbitrary software). The strategy mirrors Meta's hierarchical agent approach from two days ago, where specialized modules compensate for core agent limitations. However, Qwen-CUA doesn't address the validation gap that OpenAI's research modernization work exposed in August: agents generate plausible but incorrect outputs that evade detection. Operating on screenshots makes that problem worse, not better, since visual feedback is easier to hallucinate than structured API responses. The SWE-Touch benchmark also highlighted that real agent deployment involves human intervention mid-task, yet Qwen-CUA's training doesn't appear to stress-test collaborative scenarios.
If Alibaba publishes independent benchmarks on the GPQA Diamond or other out-of-distribution reasoning splits within 60 days, that tests whether the 397B parameter model actually reasons or merely pattern-matches on training-adjacent tasks. More critically, watch whether Alibaba releases a public instance where users can attempt real software tasks (not curated demos) and whether error rates remain below 5% on tasks involving financial or medical software, where hallucinated actions carry real cost.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsAlibaba · Qwen · Qwen-CUA · Qwen mixture-of-experts
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Qwen-CUA: Native Computer Use for (almost) Everything”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.