mage-vl-cowork: A 4B Agent That Clicks and Codes

The most interesting agent release of the past week is mage-vl-cowork, Microsoft's open-weight 4B computer-use agent, and the interesting part is not just the benchmark score. The model drives a desktop two ways at once: it clicks and types on what the screen shows, and it runs code in a persistent terminal, with a single policy deciding at each turn which channel fits. The team reports 66.0% on OSWorld-Verified under a 50-step budget, the best score they cite at the 4B size, on a benchmark where human performance is 72.4%. The weights are Apache-2.0 on Hugging Face, and the technical report reads as much like a training-infrastructure paper as a model card. This post works through what shipped, where the numbers come from, and what matters if you are evaluating small agents for real desktop work. Source

What actually shipped

Mage-VL-Cowork is a 4-billion-parameter vision-language policy, built on a Mage-ViT visual encoder and a Qwen3-4B language backbone, released by the Microsoft Mage Team with an arXiv technical report dated October 29, 2026. It exposes two action channels: computer_use, eleven screenshot-grounded GUI primitives in normalized coordinates, and bash, a persistent shell and Python environment. The routing between them is internal: the model emits a three-part turn with a rationale, an action summary, and a tool call, and no external orchestrator or multi-agent router sits above it. The working memory is deliberately tight, an asymmetric sliding window holding the three most recent screenshots, four full responses, and a folded action history inside a 32,768-token active context. Source

The headline numbers, with their frame

The 66.0% OSWorld-Verified figure comes from the team's own evaluation under a 50-step budget, so treat it as self-reported until independent runs appear, the same caution we apply to every vendor benchmark. Inside their table, the model leads its size class: the nearest cited open agents are ScaleCUA-Qwen3.5-9B at 68.7% with more than double the parameters, and larger models like EvoCUA-32B at 56.7% and OpenCUA-72B at 44.9%, while Fara1.5-4B, the cited 4B comparison, lands at 13.9%. Frontier models still lead overall: the table shows GPT-5.5 at 78.7%, Claude Opus 4.7 at 78.0%, and human performance at 72.4%. Category scores show where the model is strong and weak: 87.5% on operating-system tasks, 64.9% office, 65.0% daily, 66.3% professional, and 62.7% on multi-application workflows, the hardest category and the one closest to real knowledge work. Source

The training recipe is the real release

The report's argument is that hybrid computer use is a training problem, not a prompting one. The five-stage specialize-then-unify curriculum starts from an intermediate Mage-VL checkpoint and builds each channel separately, GUI-oriented midtraining, GUI agent SFT, then coding and terminal SFT, before Hybrid CUA SFT teaches one policy to combine them and online hybrid RL sharpens the result. The OSWorld-Verified progression across those last three stages is 45.56% to 55.84% to 66.0%. The RL stage deserves attention beyond accuracy: it added 10.16 percentage points over Hybrid SFT while cutting mean model calls per task from 21.57 to 18.77 and output tokens from 7,374 to 5,499, a 25.4% reduction. An agent that gets better and cheaper to run in the same step is the direction every team fine-tuning agents wants confirmed. Source

Two upstream choices explain much of the gain. First, the team evolved the teacher harness before collecting demonstrations: pruning redundant tools and pre-installing domain libraries lifted the GPT-5.4 teacher from 89.6% to 92.0% with 29.2% fewer steps, and the effect amplified in the student, 37.6% versus 25.1% on held-out multi-application tasks, a 12.5-point gap from harness quality alone. Second, Mage-CUA-Forge replaced per-task agent loops with reusable, skill-conditioned generation workflows, synthesizing 293,806 tasks across 6,915 skills and retaining 318,837 verified trajectories while cutting recurring generation cost by 82.6%, from $0.906 to $0.158 per task. For anyone building agent training data, that cost curve may be the most quotable number in the report. Source

Why a hybrid channel matters for desktop agents

Pure GUI agents pay a token tax: a single screenshot costs thousands of tokens, and editing a spreadsheet cell by simulated mouse is slow and fragile. Pure code agents lose the plot when the task lives in a rendered canvas or an application with no API. The release frames hybrid as the practical middle, and the channel split in the retained training trajectories shows what the policy learned to prefer: 68.19% GUI tool calls, 31.81% terminal calls, with the model switching to bash for structured edits in documents, spreadsheets, and presentations and to computer_use for visual layouts and live application state. Notably, adding a third accessibility-tree channel made things worse in their ablation, the teacher dropped and the student fell 12.5 points, which supports a design rule we have seen elsewhere: more tools are not always better, and a clean action space beats a wide one. Source

The deployment profile

The model card makes the footprint concrete: a single unified checkpoint of about 9.48 GB in bfloat16 across two safetensors shards, using a llava_onevision2 architecture with image and text in, text out. That size fits on one modest GPU, and the architecture supports up to 262,144 context tokens even though the active training and inference window is 32,768. Training compute is disclosed at 7.79 ZFLOPs of nominal LM compute across the five stages. For teams watching the small-model line, this release extends what Microsoft's Mage family has been building toward: Mage-ViT as the visual encoder, Mage-VL as the 4B multimodal foundation, and now Mage-VL-Cowork as the agent layer on top, all published under the same license. Source

The caveats before you plan around it

OSWorld-Verified is a controlled benchmark with resettable virtual machines, not your production desktop. The 66.0% figure sits below the human reference of 72.4% and well below frontier-model scores, so this is not a replacement for a frontier-driven computer-use setup where reliability is paramount; it is a claim that the capability floor for on-device and cost-sensitive deployments has moved. The evaluation is the authors' own, the step budgets in their comparison table are uneven across models, and a 4B policy will inherit the failure modes of its Qwen3-4B backbone on long-horizon reasoning. The report also discloses serious training infrastructure, 32 NVIDIA B200 GPUs plus 1,120 CPU sandbox slots for RL, which matters if you plan to reproduce the recipe rather than just use the weights. None of this diminishes the release; it defines which claims are benchmark claims and which are operational ones. Source

What this means for agent builders, in our assessment

This section is our read, separate from the sources. The durable lesson is the harness-first finding: the quality of the scaffold, tools, and environment around a model moved the teacher four points and the student twelve, which mirrors what we see when teams tune agent setups, as covered in our managed deep agents post. The second lesson is that small open agents are becoming viable for scoped desktop tasks, which changes the build-versus-buy math for internal automation, a theme we tracked across our kimi k3.1 pro and gpt-6.1 terra coverage. If you evaluate mage-vl-cowork, do it the way the report earned its own numbers: fixed task sets, verifiable rewards, measured cost per task. Our AI agent regression testing post covers how to stand up those gates before you let any agent, 4B or frontier, act unsupervised. Source