Microsoft releases Mage-VL-Cowork, a 4B open computer-use model scoring 66.0% on OSWorld-Verified
READ THE ORIGINAL SOURCEOpen original sourcehuggingface.co · Opens in a new tabMicrosoft's Mage-VL-Cowork is a 4B vision-language policy that combines screenshot-grounded GUI interaction (computer_use) and persistent shell/Python execution (bash) in a single model without an external orchestrator. Trained from an intermediate Mage-VL checkpoint through a five-stage curriculum, the model reports 66.0% on OSWorld-Verified under a 50-step budget. The text also describes evolved teacher harnesses, a workflow-first task synthesis engine (Mage-CUA-Forge) generating 293,806 tasks, and online RL that adds 10.16 points over Hybrid SFT while reducing model calls and output tokens.
⚑ Review needed · why this is flagged
The automatic check found a question that still needs a human reading of the original source. The discovery rationale is not a verified finding.
The signal remains in the tracker so you can inspect its original source and discovery trail. This flag lowers its source evidence score; it does not judge whether the claim is true.
One primary source; other links are context, not independent corroboration.
- Primary sourcePrimary source discovery · source check open
Grok Bot submitted this discovery. The automatic check found a question that still needs a human reading of the original source. The discovery rationale is not a verified finding. The original source and submitted evidence remain inspectable; no editor has verified the claim.
Open full article view →- Original repository+40
- Feed text or abstract+15
- No editor link review+0
- Original non-discussion link+5
- Review needed-8
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
Microsoft's Mage-VL-Cowork is a 4B vision-language policy that combines screenshot-grounded GUI interaction (computer_use) and persistent shell/Python execution (bash) in a single model without an external orchestrator. Trained from an intermediate Mage-VL checkpoint through a five-stage curriculum, the model reports 66.0% on OSWorld-Verified under a 50-step budget. The text also describes evolved teacher harnesses, a workflow-first task synthesis engine (Mage-CUA-Forge) generating 293,806 tasks, and online RL that adds 10.16 points over Hybrid SFT while reducing model calls and output tokens.
A READER'S EVIDENCE CHECKHow to read this signalSix questions about claims, sources, risks, and open checks
This is a signal-level guide until claim-level evidence is available. “Not recorded” means the evidence has not been assessed here; it does not mean a result is false. How we check AI news →
- 01Who is making each claim?
The claimant and whether each claim was measured or asserted are not recorded in this public brief.
- 02What was actually tested?
Not recorded per claim. Check the source for sample size, tasks, model and benchmark versions, and reproducibility.
- 03Where are the sources?
Original link: huggingface.co. Brief basis: Hugging Face model card. Primary page date and quote locations are not established in this panel.
- 04What could mislead?
The automatic check found a question that still needs a human reading of the original source. The discovery rationale is not a verified finding.
- 05What is at risk if wrong?
Not assessed per claim. Consider safety, privacy, cost, and adoption decisions before relying on the result.
- 06What would change the view?
No claim-specific check is recorded. An inspectable primary method or independent replication could change the assessment.
Key findings from the stored brief
- Mage-VL-Cowork reports 66.0% on OSWorld-Verified under a 50-step budget, described as best among 4B models.
- The model routes turns between computer_use GUI primitives and bash, with no external multi-agent router.
- Online RL adds +10.16 percentage points over Hybrid SFT while cutting mean model calls by 13.0%.
- Evolving the teacher harness lifted the GPT-5.4 teacher to 92.0% and yielded a +12.5 pp student gain.
Source claims not independently checked; benchmark numbers are self-reported in the model card.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Grok Bot submitted this discovery. The automatic check found a question that still needs a human reading of the original source. The discovery rationale is not a verified finding. The original source and submitted evidence remain inspectable; no editor has verified the claim.
These are Grok Bot discovery notes. AINewsIQ prepared a brief from available source material. The bot's rationale, linked posts, authors, and stances have not been independently verified.
Microsoft posted Mage-VL-Cowork on Hugging Face on Oct 30 under Apache-2.0: a 4B model that switches between GUI actions and bash. Microsoft reports 66.0% on OSWorld-Verified with a 50-step budget, which it calls the best among 4B models. These are Microsoft's own numbers.
Source claims not independently checked; benchmark numbers are self-reported in the model card.
Engagement counts, when shown, are observed snapshots. They do not establish accuracy or corroboration.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from Hugging Face model card.
A publication revision history is not stored for this signal. New engagement snapshots do not mean the source itself changed.
Why this may matter
It suggests a small on-device-scale model can handle end-to-end computer use by choosing between visual GUI actions and code execution, with the released weights under Apache-2.0 allowing independent testing.
Limits and open questions
Source claims not independently checked; benchmark numbers are self-reported in the model card.
Organization-specific financial, security, and operational effects require the organization's own context and review.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.