Microsoft releases 4-billion-parameter computer-use AI
Microsoft on Oct. 30 (the same day Japan time) released the weights (the trained model itself) of "Mage-VL-Cowork," an AI model that automatically carries out computer work, on the AI model sharing site Hugging Face. A single model chooses, step by step, between looking at the screen to click and type and running commands or programs in a terminal. It is small, at about 4 billion parameters (a measure of an AI's size), and its weight files come to only about 9.5GB. In the company's evaluation, it succeeded on 66.0% of tasks on the computer-use benchmark "OSWorld-Verified." The license is Apache 2.0, which permits commercial use. Microsoft has not announced it on its blog or official X account, and aitrackerbot, an X account that automatically tracks AI model releases, reported the listing on Oct. 30 (9:21 p.m. the same day Japan time).
Outperforms 30-billion-parameter-class open models in own evaluation
OSWorld-Verified has AI carry out tasks such as spreadsheets, document creation and image editing on a Linux computer and measures the share it completes. According to Microsoft, Mage-VL-Cowork succeeded on 66.0% of 361 tasks, excluding those using Google Drive, under a limit of 50 steps (the average of two runs). In the table the company presented, it outperformed the roughly 30-billion-parameter "Muse Glimmer-30B" (65.9%), Alibaba's "Qwen3.6-27B" (63.9%) and the 72-billion-parameter "OpenCUA-72B" (44.9%), and was the highest among 4-billion-parameter-class models. "Fara1.5-4B," a model of the same size the company released for browser use, scored 13.9% on this evaluation. By task type, OS settings and similar tasks scored a high 87.5%, while tasks spanning multiple apps were the lowest at 62.7%.
The human baseline in the same table is 72.4%.
Falls short of top commercial models and the human baseline
In the same table, it trails OpenAI's "GPT-5.5" (78.7%) and Anthropic's "Claude Opus 4.7" (78.0%) by more than 12 points. It is below the human baseline (72.4%) and also falls short of the 9-billion-parameter research model "ScaleCUA-Qwen3.5-9B" (68.7%). The commercial models in the table are from generations before each company's current top model, and according to Yutori, which compiles leaderboards for the evaluation, Anthropic's "Claude Fable 5" reached 86.0% under a 100-step limit. Mage-VL-Cowork's figures are Microsoft's own evaluation and do not appear on the official leaderboard. The evaluation covers only a Linux environment, and the only language the model card lists as supported is English.
On "WebVoyager-Static," a separate evaluation of tasks on websites, the version before reinforcement learning was added scored 58.6%, below Fara1.5-4B for browser use (68.0%). It outperformed Alibaba's general-purpose models "Qwen3-VL-8B-Thinking" (53.7%) and "Qwen3.5-9B" (53.3%).
Switching between screen actions and commands to cut output
Mage-VL-Cowork is based on "Mage-VL," a model for understanding images and video that Microsoft released in July. It uses Alibaba's open model "Qwen3-4B" for the text-handling part and "Mage-ViT," trained from scratch in-house, for the image-reading part. It can choose within the same model between 11 types of actions, such as clicking and typing on screen, and running commands in a terminal. Microsoft describes it as a model that can judge when to click and when to write code. Through final-stage reinforcement learning (learning by trial and error), the number of model calls per task fell from 21.57 to 18.77, and output tokens (units of text) fell from 7,374 to 5,499. In the company's measurement, compared with OpenAI's "GPT-5.4-mini," the number of calls was nearly the same at 18.77 versus 18.24, while output tokens were less than half at 5,499 versus 12,397.
Training used 293,806 tasks synthesized by "Mage-CUA-Forge," a system that automatically creates practice tasks. According to Microsoft, the system lowered the cost of generating data per task from $0.906 to $0.158. Training used 32 of NVIDIA's "B200" graphics processing units (GPUs). The evaluation code was released, but the task synthesis system and training data were not.
Terms of use: Apache 2.0, intended for isolated environments
The weights are released under the Apache 2.0 license, allowing commercial use and modification. The amount it can read at once (context) is 32,768 tokens. Because of its custom architecture, loading it requires a setting that runs the developer's code. Because the model can execute arbitrary commands, Microsoft asked on the model card that it be run in a disposable, isolated environment and kept away from users' files and credentials. It recommends inserting human confirmation for irreversible actions such as transferring money or logging in, and restricting external communication to prevent attacks in which instructions planted in web pages hijack the AI. Downloads over the past month totaled only 752 (as of Nov. 2).
Separate from "Copilot Cowork," with shift to in-house models in focus
The similarly named "Copilot Cowork" is a paid agent to which Microsoft 365 users delegate work, and it became generally available on June 16 (U.S. time). According to research firm Directions on Microsoft, it uses Anthropic's "Claude Opus 4.8" and "Claude Sonnet 4.6" as its brains, and OpenAI's "GPT-5.5" can also be selected in an early-access program. Pricing is the "Microsoft 365 Copilot" fee ($30 per user per month) plus usage-based charges. In the same month of June, Microsoft said it would soon introduce its own model, "Cowork 1," which would handle tasks at significantly lower cost. Microsoft has said nothing about the relationship between Mage-VL-Cowork and Cowork 1.
Microsoft's July-September results, announced on Oct. 28 (Oct. 29 Japan time), showed revenue of $92.28 billion (up 19% from a year earlier), with its Azure cloud up 48%. Capital spending swelled 103% to $39.37 billion, making the payoff from AI investment a focus for the market.
