Primary use stated or inferred cautiously from official documentation.
M
MICROSOFT / microsoft/Mage-VL-Cowork
Mage-VL-Cowork
A 4B-parameter visual model from Microsoft that switches between graphical interaction and code or terminal use. It accepts text and screenshots and generates text containing actions for an external executor. The checkpoint was published on October 30.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | microsoft/Mage-VL-Cowork |
|---|---|
| Model type | Vision and documents · Code · Multimodal |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted |
| Release | 30 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image |
| Exit | Text |
| Context window | 32,768 tokens in reference training and inference |
| maximum output | Not published |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Turns with reasoning, summary and tool calls |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · SGLang · Transformers |
Research desktop agents and document editing in isolated environments with limited permissions and result verification.
The application executes the actions, not the checkpoint alone. The architecture’s 262,144 tokens do not establish quality at that length: the published active window is 32,768. The SFT-stage WebVoyager-Static result is not attributed to the final checkpoint.
03 / CAPABILITIES
What it can do
Orientation
Research desktop agents and document editing in isolated environments with limited permissions and result verification.
Context
32,768 tokens in reference training and inference · Not published
Tools and integration
Not published
Access
Hugging Face · downloadable checkpoint
Input modalities
TextImage
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Apache 2.0 weights for self-hosting. No commercial API or price has been verified. Task-synthesis costs in the report are not inference rates; budget for GPU, memory and the execution environment.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Microsoft reports 66.0% on OSWorld-Verified: the average of two runs over 361 tasks excluding Google Drive, with 50 steps. It uses MAGEHybridAgent, 1920 × 1080 screenshots and a specific history. This is not a universal agent comparison or a test reproduced by Inferama.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| OSWorld-Verified · Mage hybrid 50 steps | 66,0 % | Task success | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Mage-VL-Cowork
Version added to Inferama's verified catalog.
07 / SOURCES