Mage-VL-Cowork Report - WebVoyager-Static: leaderboard
Microsoft Mage-VL-Cowork technical report Table 4, own offline reruns of all 643 tasks across 15 websites under identical offline settings. Deterministic verifier checks required page visits/UI operations, answer and backend database state. Percent successful tasks (Mage 377/643=58.6%). Self-reported live comparator percentages and score deltas are separate protocols and not ingested. Evaluated Mage checkpoint is SFT, not finalRL. QwenThinking checkpoints explicit; no unspecified effort inferred. Source does not give fixed web step budget. Metric: verified success percent.
Metric: Verified task success (%). Source: raw.githubusercontent.com. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B (Thinking) | 53.7 | #423 |
| 2 | Qwen 3.5 9B | 53.3 | #391 |
| 3 | Qwen 3 VL 4B (Thinking) | 49.6 | #548 |
| 4 | Qwen 3.5 4B | 37.3 | #514 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mage-vl-cowork-report-webvoyager-static · How It Works · Data refreshed daily, snapshot 2026-11-05.