Microsoft Releases Mage-VL-Cowork: An Open 4B Agent That Clicks *and* Codes on Your Desktop
Primary sources: microsoft/Mage-VL-Cowork model card on Hugging Face (repo created Oct 30, 2026, 8:13 AM ET per the Hugging Face API), Mage-VL-Cowork project page, code on GitHub. Checked Oct 30, 2026 ~8:10 PM ET.
Oct 30, 2026: Most computer-use AI agents work like a very patient intern. They take a screenshot, click, take another screenshot, and click again. Microsoft's new Mage-VL-Cowork takes a different approach. It's a 4-billion-parameter open model that decides on each turn whether to click through the interface or drop into a terminal and write code. Microsoft says it scores 66.0% on OSWorld-Verified, and the model card calls that "the best performance among 4B models."
The weights are on Hugging Face under the Apache-2.0 license.
What it is
According to the model card, Mage-VL-Cowork is a single vision-language model with two tools:
computer_use: clicks, drags, typing, key presses and scrolling, guided by screenshotsbash: a persistent shell and Python environment for structured edits and checks
Microsoft's tagline: "One compact 4B policy that knows when to click — and when to code." The card says it routes between the two channels by itself, "with no external orchestrator." Its own example shows the model getting a spreadsheet task, deciding that editing the .xlsx file with Python is "much faster and more reliable than opening Calc and clicking through cell menus," and doing that instead.
Key specs from the card:
| Spec | Value (per Microsoft's model card) |
|---|---|
| Parameters | 4B (9.48 GB in bfloat16) |
| Backbone | Mage-ViT vision encoder + Qwen3-4B (Qwen3-4B-Instruct-2507) |
| Context | 32,768 tokens used in training and inference (architecture supports up to 262,144) |
| Memory | Keeps the 3 most recent screenshots + 4 full responses; older steps folded into one-line summaries |
| Training compute | 7.79 ZFLOPs across 5 stages, on 32 NVIDIA B200 GPUs |
| License | Apache-2.0 |
The benchmark numbers (self-reported)
All figures below are Microsoft's own. We haven't seen independent replication yet.
- OSWorld-Verified: 66.0%, the average of two full runs (65.33% and 66.75%) on 361 tasks (Google Drive tasks excluded) with a 50-step budget
- 87.5% on the benchmark's operating-system tasks and 62.7% on multi-application workflows
- Online reinforcement learning added +10.16 points over the previous training stage while cutting average model calls by 13.0% and output tokens by 25.4%
The card's own table adds some useful context:
- Frontier closed models still score higher in that table: GPT-5.5 at 78.7 and Claude Opus 4.7 at 78.0. The human reference is listed at 72.4.
- An open 9B model, ScaleCUA-Qwen3.5-9B, scores 68.7, above Mage-VL-Cowork, at more than twice the size.
- Among models the card labels 4B, the other entry, Fara1.5-4B, is listed at 13.9.
So "best at 4B" is the right way to read the claim. It isn't "best overall."
How Microsoft says it trained it
The card describes a five-stage "specialize-then-unify" curriculum:
- GUI-oriented midtraining (visual grounding and tool-call formatting)
- GUI agent fine-tuning on multi-step desktop and web tasks
- Coding and terminal fine-tuning
- Hybrid fine-tuning that teaches the model when to switch channels
- Online reinforcement learning with verifiable task rewards
To make the training data, Microsoft built a synthesis engine called Mage-CUA-Forge. The card says it generated 293,806 tasks across 6,915 atomic skills. It says the engine cut recurring per-task generation cost by 82.6% (from $0.906 to $0.158 per task) compared with building each task through a standard agent loop.
Safety: run it in a sandbox
Microsoft is direct about the risk. The card says the model "is an autonomous hybrid agent capable of executing arbitrary shell commands," and that it was "designed and evaluated for isolated, disposable sandbox environments." Its recommendations:
- Run it only inside an isolated container or VM
- Require human confirmation before irreversible actions such as payments, logins or network submissions
- Don't mount SSH keys, cloud tokens, browser cookies or password stores in its workspace
- Restrict network access to reduce prompt-injection risk from web pages and documents
Why it matters
A 4B model is small enough that local or on-device computer-use agents become realistic. The card explicitly frames the release as "pushing the frontier of on-device intelligence." The bigger lesson may be the design, though: letting the agent write code instead of clicking whenever that's more reliable.
What's confirmed / not confirmed
Confirmed (Microsoft's model card and Hugging Face repo): - Open weights on Hugging Face under Apache-2.0; repo created Oct 30, 2026 - 4B parameters, Qwen3-4B language backbone, GUI + bash tool channels - Self-reported 66.0% OSWorld-Verified (two-run average, 50 steps, 361 tasks) - Evaluation code and an SGLang Docker launcher are included
Not confirmed: - Independent reproduction of the 66.0% score - How well it runs on consumer hardware. The card gives the bf16 size (9.48 GB) but no consumer-GPU or laptop benchmarks. - Whether Microsoft will use it in any product. The card doesn't say. - The technical report is linked as a PDF on GitHub; we didn't see an arXiv ID.
Related
Up: AI
