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 hybrid computer-use agent, on Hugging Face under the Apache-2.0 license. The model card reports 66.0% on OSWorld-Verified under a 50-step budget, which Microsoft says is the best result among 4B models.

What it is

Mage-VL-Cowork is a single vision-language policy (image + text in, text out) that handles screenshot-grounded GUI interaction (computer_use) and persistent code/terminal execution (bash) with no external orchestrator. It was trained from an intermediate 4B Mage-VL checkpoint through a five-stage specialize-then-unify curriculum.

Specifications

• Developer: Microsoft Mage Team
• Parameters: 4B (9.48 GB in bfloat16 across 2 safetensors shards)
• Backbone: Mage-ViT (16×16 patch grid, spatial_merge_size=2, 3D RoPE) plus Qwen3-4B (Qwen3-4B-Instruct-2507)
• Architecture: llava_onevision2
• Context length: 32,768 tokens for active training and inference; the architecture supports up to 262,144
• Working memory: asymmetric sliding window of the 3 most recent screenshots, 4 full responses and a folded Action: history
• Action channels: computer_use (11 GUI primitives in [0, 999] normalized coordinates) and bash (persistent shell and Python environment)
• Output format: a <think> rationale, an Action: summary and a <tool_call>
• Training compute: 7.79 ZFLOPs nominal LM compute across 5 stages, on 32 NVIDIA B200 GPUs plus 1,120 CPU sandbox slots for RL
• Pricing: not yet disclosed (open weights)

Benchmark results

According to Microsoft, the model scores 66.0% on OSWorld-Verified, with 87.5% on Operating System tasks and 62.7% on Multi-Application Workflows. Standalone grounding improved from 30.6% to 56.9% on OSWorld-G and from 78.8% to 87.2% on ScreenSpot-v2. Across the hybrid stages, accuracy moved from 45.56% to 55.84% to 66.0%.

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

Training pipeline

Microsoft describes two upstream components:

• Evolved teacher harness: the team pruned overlapping tools such as xdotool and accessibility trees and pre-installed domain Python libraries. This lifts the GPT-5.4 teacher to 92.0% with 29.2% fewer steps, and the student gain at 4B is +12.5 pp (37.6% vs. 25.1%).
• Mage-CUA-Forge: a workflow-first synthesis engine that generates 293,806 tasks across 6,915 atomic skills and retains 318,837 verified hybrid trajectories (68.19% GUI tool calls, 31.81% terminal tool calls). Microsoft says it cuts recurring per-task generation cost by 82.6%, reducing token use by 93.3% and expenditure from $0.906 to $0.158 per task.

The curriculum starts with GUI-oriented midtraining (24.6B tokens per epoch), then GUI agent SFT (156.7K trajectories), then coding and terminal SFT, before unification through Hybrid SFT and asynchronous Online Hybrid RL.

What this means

The results address a specific cost in computer-use agents: screenshots consume thousands of tokens, and routine file edits done through clicks are slow. By learning to route between GUI and bash inside one 4B policy, Mage-VL-Cowork shows that channel selection can be trained into a small model rather than handled by a separate router. The reported gains from harness pruning (+12.5 pp for the student) indicate that teacher scaffold design affects distillation quality, not only teacher scores. The Apache-2.0 license and 4B size make the checkpoint practical to run and test locally, though the 66.0% figure is Microsoft's own report and applies to a 50-step budget.