Is Mage-VL-Cowork state of the art?Pending benchmarks.Announced Oct 30, 2026; no independent results yet.
Microsoft's 4B open-weight computer-use agent. One vision-language policy works a desktop either through screenshot-grounded GUI actions or through a persistent bash and Python shell, choosing the channel turn by turn. Fine-tuned from an intermediate Mage-VL checkpoint; Apache-2.0.
The details so far
- Lab
- Microsoft
- Announced
- Oct 30, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 4.74B stored (BF16 safetensors); Microsoft lists it as 4B
- Type
- Hybrid computer-use agent: image + text in, text out (think rationale, action summary and a tool call)
- Base model
- microsoft/Mage-VL (intermediate 4B checkpoint)
- Backbone
- Mage-ViT vision encoder + Qwen3-4B-Instruct-2507 (llava_onevision2 architecture)
- Action channels
- computer_use (11 GUI primitives on normalized [0, 999] coordinates) and bash (persistent shell and Python environment)
- Context
- 32,768 tokens used in training and inference; architecture supports up to 262,144
- Training
- Five stages: GUI midtraining, GUI agent SFT, coding and terminal SFT, hybrid SFT, then online GRPO on 1,988 verifiable tasks; 32 NVIDIA B200 GPUs
- License
- Apache-2.0
- Intended use
- Released for research purposes, not for product or service deployment (Mage repo notice)
- Release
- Weights on Hugging Face Oct 30, 2026; project page and README were committed Oct 27–28 without weights
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 66.0% on OSWorld-Verified (two-run average over 361 tasks excluding Google Drive, 50-step budget, Microsoft's MAGEHybridAgent harness), with 87.5% on OS tasks and 62.7% on multi-application workflows.
Lab claim: online RL adds 10.16 points over the hybrid SFT stage while cutting model calls per task by 13.0% and output tokens by 25.4%.
Sources
REPOmicrosoft/Mage-VL-Cowork model card (Hugging Face)huggingface.co/microsoft/Mage-VL-Cowork · Oct 30, 2026LABMage-VL-Cowork project pagemicrosoft.github.io/Mage/cowork/LABMage-VL-Cowork technical report (PDF)github.com/microsoft/Mage/blob/main/assets/mage_vl_cowork_tech_report.pdf · Oct 30, 2026REPOMage-VL-Cowork code (GitHub)github.com/microsoft/Mage/tree/main/mage_vl_coworkTracking continues until an independent board lists it.