Microsoft releases Mage-VL-Cowork, a 4B computer-use agent scoring 66.0% on OSWorld-Verified

Microsoft has released Mage-VL-Cowork, a 4B-parameter vision-language agent that reaches 66.0% on OSWorld-Verified under a 50-step budget. Microsoft claims this is the best result among 4B models. The weights are on Hugging Face under an Apache-2.0 license.

What the model does

Mage-VL-Cowork is a single checkpoint that handles visual grounding, multi-step GUI manipulation, programmatic file and system operations, and switching between GUI and command line. It exposes two action channels: computer_use (11 GUI primitives in [0, 999] normalized coordinates) and bash (a persistent shell and Python environment). According to Microsoft, the policy picks the channel itself, with no external orchestrator or multi-agent router.

Each turn produces a three-part structured output: a <think> rationale, an Action: summary, and a <tool_call>.

Specifications

Property	Value
Developer	Microsoft Mage Team
Parameters	4B (9.48 GB in bfloat16, 2 safetensors shards)
Backbone	Mage-ViT + Qwen3-4B (Qwen3-4B-Instruct-2507)
Context length	32,768 tokens active for training and inference; architecture supports up to 262,144
Working memory	3 most recent screenshots, 4 full responses, folded Action history
Training compute	7.79 ZFLOPs nominal LM compute across 5 stages
Hardware	32 NVIDIA B200 plus 1,120 CPU sandbox slots for RL
License	Apache-2.0

Pricing is not applicable to the open weights, and hosted API pricing is not yet disclosed.

Benchmark results

Microsoft reports the following:

• OSWorld-Verified: 66.0% (50-step budget), including 87.5% on Operating System tasks and 62.7% on Multi-Application Workflows.
• OSWorld-G: 30.6% to 56.9% across the grounding stage.
• ScreenSpot-v2: 78.8% to 87.2% across the grounding stage.
• Training progression: 45.56%, then 55.84%, then 66.0% across Hybrid SFT and online hybrid RL.

Online GRPO with verifiable task rewards adds +10.16 percentage points over Hybrid SFT. Mean model calls fall from 21.57 to 18.77 (-13.0%), and output tokens fall from 7,374 to 5,499 per task (-25.4%).

Training recipe

The model starts from an intermediate 4B Mage-VL checkpoint and goes through a five-stage specialize-then-unify curriculum. The first stages build standalone GUI grounding, GUI execution and coding/terminal skills. Hybrid SFT and asynchronous online hybrid RL then unify them.

Two pieces of upstream work stand out:

• Evolved teacher harness. Microsoft pruned overlapping tools (xdotool, accessibility trees) and pre-installed domain Python libraries such as python-pptx, openpyxl, pandas, pypdf and pillow. According to Microsoft, this lifts the GPT-5.4 teacher to 92.0% with 29.2% fewer steps. Student accuracy at 4B rises from 25.1% to 37.6% (+12.5 pp).
• Mage-CUA-Forge. This workflow-first synthesis engine generated 293,806 tasks across 6,915 atomic skills, of which 38.24% are multi-application. It retained 318,837 verified hybrid trajectories (68.19% GUI tool calls, 31.81% terminal tool calls). Microsoft says recurring per-task generation cost fell 82.6%, token use fell 93.3%, and expenditure dropped from $0.906 to $0.158 per task.

What this means

The reported gains come mainly from data and scaffolding rather than model size. Trimming the teacher's tool set before rollout is credited with a 12.5 pp student improvement, and the card says frontier teachers with naive GUI plus terminal access pick suboptimal channels that students then inherit.

The efficiency numbers matter for deployment. A 9.48 GB checkpoint with an Apache-2.0 license can be inspected and fine-tuned by outside teams. The reduction in model calls and output tokens after RL points to cheaper per-task inference. All benchmark figures are Microsoft's own, reported on its model card, and the OSWorld-Verified result applies to a 50-step budget.