Mage-VL-Cowork: Specs, License & How to Run

Mage-VL-Cowork is Microsoft’s hybrid computer-use checkpoint: one ~4B-class vision-language policy that picks a GUI channel (computer_use) or a code/terminal channel (bash) without an external orchestrator. Official hosts we opened on 2026-11-01: the ungated card microsoft/Mage-VL-Cowork (created 2026-10-30T12:13:24Z; last modified 2026-10-30T12:14:43Z; 93 likes / 233 downloads), the LICENSE file, config.json, the foundation sibling microsoft/Mage-VL (445 likes / 7,544 downloads), the Mage collection, github.com/microsoft/Mage (1,736 stars / 176 forks), and the project page microsoft.github.io/Mage/cowork. API gated was false.

This identification page hangs off ChatGPT alternatives, closed LLM APIs, and Claude Code alternatives. Those mappings are editorial for a local computer-use VLM. We have no sourced quality comparison against ChatGPT or Claude Code.

It is not Iris-3B (3B pixel-space text-to-image). Generation facts stay on Iris-3B vs Qwen-Image-2.1-Turbo. A same-class on-device VLM already here is MiniCPM-V 4.7 — MiniCPM-V 4.7 vs Iris-3B (VLM vs T2I, not a CUA bake-off).

The card’s “best performance among 4B models” and “on-device intelligence” sentences stay vendor.

What we verified (2026-11-01)

Field	Value	Source
Host	huggingface.co/microsoft/Mage-VL-Cowork (not gated)	HF API
Created / last modified	2026-10-30T12:13:24Z / 2026-10-30T12:14:43Z	same
Likes / downloads	93 / 233	same
License tag + file	apache-2.0 + Apache License 2.0	HF cardData + LICENSE
Pipeline tag	image-text-to-text	HF API
Architecture class	llava_onevision2	HF config
safetensors total	4,741,793,792 BF16 (~4.74B)	HF API
usedStorage	9,495,637,963 bytes (~9.50 GB)	same
Context (card / config)	32,768 active / 262,144 max_position_embeddings	README + config.json
GitHub stars / forks	1,736 / 176	GitHub API microsoft/Mage
GitHub license API	MIT	same

Card overview (copied): trained from an intermediate 4B Mage-VL checkpoint through a five-stage specialize-then-unify curriculum; action channels computer_use (11 GUI primitives in [0, 999] normalized coordinates) and bash (persistent shell and Python); working memory of the 3 most recent screenshots plus 4 full responses. Tags include computer-use, CUA, GUI-agent, coding-agent, terminal-agent.

How to run (from the official card, not tested)

Copied names only. forkalt has not loaded a checkpoint and has not run a computer-use sandbox.

The card points at github.com/microsoft/Mage (mage_vl_cowork/) and the project page. Hugging Face library_name is transformers with custom code (configuration_llava_onevision2 / modeling_llava_onevision2). The card’s safety notes require an isolated container or VM, scoped authorization, and no host credentials in the agent workspace. Those are upstream operational rules, not a forkalt runbook.

Vendor benches (self-reported)

Copied card rows only: OSWorld-Verified 66.0% under a 50-step budget; online RL +10.16 points over Hybrid SFT; OSWorld-G 30.6% → 56.9%; ScreenSpot-v2 78.8% → 87.2%. Those are self-reported. They are not a ChatGPT, Claude Code, or Nemotron 3.5 Bolt ranking.

License

LICENSE is Apache License 2.0. The GitHub repo microsoft/Mage is tagged MIT on the license API. We record both. Not legal advice.

What we did not invent

• Peak VRAM and tok/s are not yet verified.
• We do not claim visual or agent quality versus ChatGPT, Claude Code, MiniCPM-V 4.7, or Iris-3B.
• Mage-VL (foundation) is a different card; this page identifies Mage-VL-Cowork.

Related: MiniCPM-V 4.7 — WeVisOne-VL (Tencent perception VLM, different job) — Nemotron 3.5 Bolt (74B/9B local-agent VLM) — Spark-X2.6-4B (text-only ~4B agent) — Cline / OpenCursor on Claude Code alternatives. Image generation is a different search: Iris-3B — Iris-3B vs Qwen-Image-2.1-Turbo. Serving: vLLM. Catalog: home. Corrections: contact.