Almost every model that drives a computer does the same thing: look at a screenshot, compute a coordinate, click. Microsoft's Mage team open-sourced a model that asks a blunt question — how much tax does a pixel-only agent pay? A single 1920×1080 screenshot burns thousands of tokens, and re-deriving one structured spreadsheet edit through dozens of mouse drags throws away the fastest, most verifiable channel on the machine: writing code.

Mage-VL-Cowork is that bet. A 4B-parameter model built on Mage-ViT (16×16 patch grid, 3D RoPE) plus Qwen3-4B, released under Apache-2.0 with the full technical report on arXiv. It folds GUI and terminal into two action channels inside a single policy — no external orchestrator — and decides at every turn whether to click or to type.

The score: 66.0%, and what it actually means

The final RL checkpoint reaches 66.0% on OSWorld-Verified — 361 held-out tasks, Google Drive excluded, 50-step budget, averaged over two evaluation runs (65.33% and 66.75%). Category breakdown: OS 87.5, Office 64.9, Daily 65.0, Professional 66.3, Workflow 62.7.

In the paper's own comparison table, the other 4B entry, Fara1.5-4B, manages 13.9%. Going up in size, Muse Glimmer-30B scores 65.9%, Qwen3.6-27B 63.9%, EvoCUA-32B 56.7%, OpenCUA-72B 44.9%. Among open models only ScaleCUA-Qwen3.5-9B (68.7%) edges past it. On the closed side, the human reference line sits at 72.4%, GPT-5.5 at 78.7%, and GPT-5.4-mini at 69.6%.

The paper's "state-of-the-art at its parameter size" claim therefore means one specific thing: nobody else occupies the 4B tier, not that it leads everywhere. Beating the 30B model costs 0.1 percentage point; beating the 72B model costs 21. One caveat worth keeping: the table was run by the paper's own authors, and the competing scores all come from that same table rather than independent replications.

The valuable part is not the score, it is "fix the teacher first"

The clearest evidence is a harness ablation. With the teacher model pinned to GPT-5.4, on 672 held-out multi-application tasks, a pruned GUI-plus-code harness teacher scores 92.0% at 25.0 mean steps; the variant with an extra accessibility channel drops to 89.6% at 35.3 steps. Train 4B students on each teacher's trajectories and the gap widens to 37.6% versus 25.1% — a 12.5 percentage-point swing, with trajectories 21.5% shorter.

More tools does not mean a stronger agent. A frontier teacher can absorb a redundant channel by taking extra steps; a 4B student is acutely sensitive to supervision that has been split apart.

Data synthesis: per-task cost cut to a sixth

Mage-CUA-Forge stops running an open-ended agent loop per task and instead authors one reusable, skill-conditioned workflow per atomic skill. Under the same generation model, tokens per task drop from 1,276K to 85K (−93.3%) and cost falls from 0.906 USD to 0.158 USD (−82.6%). After three rounds of student-feedback iteration the pipeline has synthesized 293,806 tasks across 6,915 atomic skills, 38.24% of them spanning multiple applications, and retains 318,837 verified hybrid trajectories — 68.19% GUI tool calls, 31.81% terminal calls.

Five stages, and an RL stage that gets cheaper as it gets better

The recipe runs: GUI-oriented midtraining → GUI agent SFT → coding and terminal SFT → hybrid CUA SFT → online hybrid RL (asynchronous GRPO rollouts with verifiable task rewards). Build each channel on its own first, then teach the model when to switch, rather than letting routing become a patch over weak GUI execution.

The score climbs along that path: 45.56% → 55.84% → 66.0%. Stage 1 alone takes OSWorld-G from 30.6% to 56.9% and ScreenSpot-v2 from 78.8% to 87.2%, passing Qwen3.5-4B (54.5%) and Qwen3-VL-4B (85.3%).

The RL stage is the counterintuitive part: accuracy rises while consumption falls. Mean model calls per task go 21.57 → 18.77 (−13.0%), mean output tokens 7,374 → 5,499 (−25.4%). Against GPT-5.4-mini — 18.24 calls and 12,397 output tokens — the 4B model matches it on interaction turns while emitting 55.6% fewer output tokens, which the team puts at 2.25× more token-efficient.

What I think

The part of this work worth copying is not the architecture. It is the ordering: fix the teacher's harness before you collect data. Two years of agent talk has assumed the model is the bottleneck. This paper turns the camera around — one redundant tool bolted onto the teacher scaffold costs a 4B student 12.5 points, and that loss is essentially invisible in the training logs.

The second layer is that "hybrid" is becoming the default shape. A pure-GUI route runs into the token wall and a reliability wall; a pure-code route cannot see the canvas. A hybrid policy hands the choice back to the model. The side-by-side demo on the project page shows the shape clearly: on the same LibreOffice Calc task — merge banner ranges, label them, add a chart — the 4B model rewrites the workbook with Python in the terminal and finishes in 39 seconds across 6 steps, while Qwen3.6-27B clicks and drags for 9 minutes 56 seconds across 39 steps. One demo is not a general claim, but the direction is clear.

On "on-device," the honest version: independent tracker ModelFit lists the only quantized build as a 2.716 GB Q4_K_M GGUF plus a 0.353 GB vision projector, so it fits on a 16GB Mac. But this thing drives a full Linux desktop, which is a long way from "an agent on your phone."

So the question is not whether a small model can catch up with a large one. It is: when an agent's capability gap is no longer set by parameter count, what is it set by — and who writes the scaffolding that decides?