Microsoft releases 4B open agent that beats 72B rivals

Summary

• Microsoft published Mage-VL-Cowork, a 4-billion-parameter computer-use agent, on Hugging Face under Apache-2.0.
• It scored 66.0% on OSWorld-Verified, above open models up to 72 billion parameters.
• Commercial licensing lets companies run an agent that clicks and executes code on their own machines.

The latest

A 4-billion-parameter agent that operates a computer scored 66.0% on the OSWorld-Verified benchmark, outperforming open models more than ten times its size. Microsoft published the weights for Mage-VL-Cowork on Hugging Face under an Apache-2.0 commercial license, part of the Mage family built by a team led by Xiaoyi Zhang with around twenty co-authors. The model runs in roughly 9.48 gigabytes at bfloat16 precision.

Details

• What it does: The model is a vision-language policy that decides, at each turn, whether to take a graphical interface action with mouse and keyboard or to execute code and terminal commands in bash and Python. The model page describes it as one 4B policy that knows when to click and when to code.
• The size question: The page lists the model at 4 billion parameters in its text and 5 billion in its metadata. Either figure places it far below the open systems it outscores, and below the frontier models still leading the benchmark.
• The benchmark breakdown: On OSWorld-Verified with a 50-step budget, the overall score was 66.0%. By category: operating system 87.5, office 64.9, daily use 65.0, professional 66.3, workflow 62.7. The operating-system result is the widest margin in the published table.
• The comparisons: Project-page figures put SCALECUA-Qwen3.5-9B at 68.7, Muse Glimmer-30B at 65.9, Qwen3.6-27B at 63.9, EvoCUA-32B at 56.7, OpenCUA-72B at 44.9 and Fara1.5-4B at 13.9, with human performance at 72.4. Proprietary leaders GPT-5.5, Claude Opus 4.7, GPT-5.4 and Kimi-K2.6 sit between 73.1 and 78.7, above the human baseline.
• How it was trained: Training started from an intermediate Mage-VL checkpoint and ran five stages: interface-oriented mid-training, supervised interface-agent tuning, coding and terminal tuning, hybrid tuning, then online reinforcement learning using GRPO with verifiable rewards.
• The gains: Scores moved from 45.56% after interface tuning to 55.84% after hybrid tuning and 66.0% after reinforcement learning, a 10.16-point jump in the final stage. Average model calls fell 13.0%, from 21.57 to 18.77, and output tokens fell 25.4%, from 7,374 to 5,499.
• The data: Developers report 293,806 desktop and web tasks across 6,915 skills, 38.24% of them multi-application, plus 318,837 verified hybrid trajectories split 68.19% interface calls and 31.81% terminal calls. Task generation cost dropped 82.6%, from $0.906 to $0.158, and tokens per task fell 93.3%.
• Running it: The active context window is 32,768 tokens, with architecture support up to 262,144. Working memory holds the last 3 screenshots and 4 full responses, with 11 primitive interface tools and a cap of 15 tool calls per response. The custom architecture requires trust_remote_code=True. Training used 32 NVIDIA B200 processors and 1,120 isolated environments.
• The safety guidance: The model card says the agent was designed and evaluated for isolated, restorable environments such as Docker containers or virtual machines, because it executes arbitrary terminal commands. Developers recommend human confirmation before irreversible actions, including external sending, logins, financial transactions and permission changes.
• Keys out of reach: The card advises keeping SSH keys, cloud tokens, cookies and password stores outside the agent workspace, and restricting network access when handling untrusted web content to reduce prompt-injection risk.

Background

Mage-VL-Cowork is one of three models in Microsoft's Mage family, alongside Mage-VL for image and video understanding and Mage-Flow for generation and editing. Code sits on GitHub under microsoft/Mage with an accompanying technical report.

Between the lines

The combination that matters for enterprises is not the score alone but the package: Apache-2.0 licensing, a 9.48-gigabyte footprint and a result above OpenCUA-72B. That makes local deployment on company hardware plausible without a frontier-model API. The safety section is the counterweight — an agent that runs arbitrary terminal commands and can be steered by untrusted web content is why the developers scope it to disposable sandboxes.

What's next

Watch whether independent reproductions of the 66.0% OSWorld-Verified result appear, whether Microsoft clarifies the 4B versus 5B discrepancy on the model page, and whether the Mage family adds further computer-use releases.

Source: Microsoft (Hugging Face), ModelFit, i-scoop