whatmodel.sh / Mage-VL-Cowork

Mage-VL-Cowork is an open-weights language model from Microsoft, released October 2026. Its published size is 4.74B parameters and a 33k-token context window. Licence: apache-2.0.

Catalogue identity notes

Microsoft Mage Team 4B hybrid computer-use agent (GUI computer_use + bash) trained from an intermediate Mage-VL checkpoint (Mage-ViT + Qwen3-4B-Instruct-2507) with a five-stage curriculum ending in online RL; image+text -> text, custom llava_onevision2 architecture needing trust_remote_code. Specialised for desktop/terminal agent use with its own system prompt and tool schema, not a general chat model. Weights released 2026-10-30 (commit 6fbe7f9 'Release Mage-VL-Cowork model weights, card, and OSWorld evaluation reference'). Only the final RL checkpoint's OSWorld-Verified figure is assigned; Stage-1 grounding scores (OSWorld-G, ScreenSpot-v2) and the WebVoyager-Static score (Mage-VL-Cowork-SFT checkpoint) belong to other checkpoints. GGUF builds are ggml-org automatic conversions. No API.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

Card states '4B (9.48 GB in bfloat16 across 2 safetensors shards)'; 4.74 is the HF safetensors metadata count (4,741,793,792 BF16 parameters at sha 6fbe7f9), which includes the Mage-ViT vision encoder. Dense (Qwen3-4B-Instruct-2507 language backbone).

No comparable score yet. Its published results are below, but they don't overlap enough with other models to calculate a score.

$ how it compares

Scored LLMs released closest to Mage-VL-Cowork, out of 1206 with a score, then more from Microsoft.

#modelscorewhereRAMreleased
3Gemini 4 Carbon Google84±1api $4.00/M2026-10
6GPT-6.1 Terra OpenAI83±1api $2.00/M2026-10
10Muse Wonder Full Meta81±1Muse Code only2026-10
19Kimi K3.1 Pro Moonshot AI78±1api price unknown2026-10
20Muse Wonder Meta78±1api $8.00/M2026-10
more from Microsoft
345Phi-4-reasoning-plus 14B Microsoft58±1local~11GB2025-04
368Phi-4-reasoning 14B Microsoft57±1api $0.22/M +local~10GB2025-04
523Phi-4 14B Microsoft53±1api $0.087/M +local~11GB2024-12

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasksnot ranked on this task yet · 1 test
testresultvs the others in its rundifficultycountswhy
OSWorld-Verified lab66.0% ✎█▃·▂▅▂▅▂▅▃▆▆ beats 14/28—not counted yet: no other model in the same run to compare it with

$ where to run it

Memory is what a version takes in GPU memory (or a Mac's unified memory) with an 8k-token conversation. With less, it runs partly from system RAM, more slowly. Enter your machine to see what runs at full speed.

Other ways to run it

version

  1. Run it as published · needs 4.7 GB memory, 3.07 GB download · from huggingface.co/ggml-org
    llama serve -hf ggml-org/Mage-VL-Cowork-GGUF

Model repository and documentation: huggingface.co/microsoft ↗

$ published results · 1 result, not enough overlap for a score
testresultsource and conditions
OSWorld-Verified66.0%microsoft/Mage-VL-Cowork model card · maker-reported

Microsoft Mage Team own evaluation in the pinned Mage-VL-Cowork card, section '1. Main Desktop Results: OSWorld-Verified', row 'Mage-VL-Cowork-4B (Ours)', column 'Overall' = 66.0; protocol text gives the unrounded 66.04% as the two-run average (65.33% run 1, 66.75% run 2). 361 OSWorld-Verified tasks excluding Google Drive, 50-step budget, MAGEHybridAgent harness (history_n=3, images_to_keep=3, 1920x1080), temperature 0.7, top_p 0.8, max_new_tokens 4096. Success rate, final Online-RL checkpoint. Other rows in that table are other models (some with 100-step budgets) and are not used.

$ improve this page

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json