Alibaba’s MAI-UI Team Open-Sources OpenGCUA for Generalist Computer-Use Agents

The MAI-UI team at Alibaba Token Hub, Alibaba Group has open-sourced OpenGCUA (Open Generalist CUA), a stack for what the authors call generalist computer-use agents — systems that mix visual GUI control with programmatic access, and let the model choose how to observe, act and organise its work. The first release is documented in an arXiv preprint submitted on 15 October 2026, with code on GitHub under Apache-2.0 and a project page.

The release itself is confirmed. The benchmark and efficiency figures below are the authors’ own claims in a preprint that has not been peer reviewed. Quote them as such.

What they mean by GCUA

The abstract defines the term directly:

We define a generalist computer-use agent (GCUA) as a system in which the model chooses how to observe, act, and organize its work across visual and programmatic interfaces.

The authors argue that frontier computer-use systems — they name GPT-6 Astra’s computer-use stack as an example — “remain largely closed, limiting systematic study of their design.” OpenGCUA is framed as an ongoing project to open the full stack (agent design, harness, SDK, trajectories, and later models and real-device benchmarks) with the community.

This first report does not release a newly trained model. It studies prompts, tool interfaces and a harness that let existing frontier models use code execution and what the authors call harness-to-model decision transfer (moving selected observation and task-state decisions from fixed harness rules to the model).

What is in the first release

According to the abstract and body, the team releases:

• Implementation and prompts
• Evaluation trajectories from all eight models tested
• Preliminary versions of a real-device harness and a cross-device SDK

The paper’s closing note: “The MAI-UI Team presents the first release of OpenGCUA; further project releases will follow.” The harness and SDK are explicitly preliminary, not finished products.

The eight models named in the evaluation include GPT-6 Astra, Claude Opus 5, Gemini 3.8 Flash, Qwen3.8-Max, Qwen3.8-Flash, Kimi K3, DeepSeek-V4.1-Flash and GLM-5.3-Flash.

Authors’ OSWorld-v2 claims (offline subset)

On the offline subset of OSWorld-v2 (82 tasks that exclude internet-dependent items, per the paper’s body), the authors say their full OpenGCUA configuration improves task scores across all eight models relative to batched GUI-only execution. Exact abstract wording:

On OSWorld-v2, our full configuration improves task scores across all eight evaluated models, including GPT-6 Astra and Claude Opus 5, raising mean partial score by +19.6 points and reducing reported steps per task by an average of 53.0% relative to batched GUI-only execution.

The body clarifies that this is the offline subset, and that the score is a partial score (mean evaluator partial credit on a 0–100 scale), not an independent third-party leaderboard. Treat the +19.6 and 53.0% figures as author-reported, on their harness and their selected runs.

Code execution beyond shell access

A central design claim is that writing GUI operations as code (with loops, branches and intermediate checks) helps more than giving the same model GUI and shell access with batched actions. The body states that, with the same GUI and shell access, writing GUI operations as code “adds 6.1 points to the mean partial score and cuts steps per task by 14.8% against the GUI+CLI baseline.”

Again: that is the authors’ ablation claim, not an external replication.

Real-device harness: efficiency gap (authors’ four macOS tasks)

The team also reports early runs under its preliminary real-device harness on four real macOS tasks. The abstract says:

Across four real macOS tasks under our real-device harness, selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra

Figure 1 in the paper adds that Qwen3.8-Max “completes fewer tasks” in that comparison. With only four tasks and selected runs, the authors themselves treat this as a characterisation of the evaluated configurations, not a general model ranking.

Caveats to keep in view

• Own benchmarks: OSWorld-v2 offline scores and ablations are the authors’ measurements under their OpenGCUA harness.
• Partial scores: the headline +19.6 figure is mean partial score, not full-task success alone (though they also report success rising for every model).
• Offline subset: not the full OSWorld-v2 suite; internet tasks are excluded.
• Preliminary harness and SDK: real-device components are early; more releases are promised.
• No peer review yet: arXiv preprint v1.
• No new weights this round: design validation on existing frontier models.

The same team’s MAI-Gym platform is described in a separate 14 October preprint (arXiv 2610.20939); that is background only and is not the subject of this post.

Why it matters

Computer-use agents are shifting from click-only GUIs toward hybrid stacks that run code, call shells and share a machine with a human user. Most of that frontier work is closed. An open harness, prompts and released trajectories — even without a new model — give outsiders something concrete to reproduce and argue with. The open question is whether the +19.6 / −53% gains hold under independent evaluation, and how far the preliminary real-device pieces go once users, overlapping windows and changing requests enter the loop.

For related TSN coverage of computer-use tooling and models, see Anchor’s Action Index for shared web workflows and Claude Haiku 5.5’s cheaper computer-use positioning.

What’s still unknown

• Independent reproduction of the OSWorld-v2 offline gains
• When a trained OpenGCUA model ships (roadmap item; not in this release)
• How the preliminary SDK behaves on phones and multi-device workflows outside the authors’ demos
• Peer-reviewed assessment of the claims

What this means

• The MAI-UI team at Alibaba Token Hub, Alibaba Group, has open-sourced OpenGCUA (Apache-2.0) with a 15 October 2026 preprint.
• Authors claim +19.6 mean partial score and 53.0% fewer steps vs batched GUI-only on the OSWorld-v2 offline subset, across eight models.
• Code-as-GUI claim: +6.1 points and −14.8% steps vs GUI+CLI baseline (authors).
• Real-device: selected Qwen3.8-Max runs use 19.4× GPT-6 Astra’s reported reasoning tokens on four macOS tasks; fig. 1 says it completes fewer tasks.
• Harness and SDK are preliminary; further releases planned; no new model weights in this drop. Results are preprint claims, not peer reviewed.

Sources

• Zhou et al. / The MAI-UI Team, “OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use”, arXiv:2610.22200, 15 October 2026
• Tongyi-MAI/OpenGCUA on GitHub (Apache-2.0)
• OpenGCUA project page

Related stories

• Anchor Launches Action Index, a Shared Memory of Web Workflows for Computer-Use Agents
• Claude Haiku 5.5: Anthropic’s Small Model Gets Much Cheaper and Much Better at Using a Computer