Mage-VL-Cowork: Microsoft's 4B computer-use model that clicks or codes
A 4-billion-parameter open-weight vision-language model that operates a desktop two ways inside one policy: clicking through screenshots, or dropping to a bash terminal when code is the faster route. Microsoft reports 66.0% on OSWorld-Verified.
Why it matters
Channel routing in a tiny model: one 4B policy decides per step whether to use the GUI or the terminal, beating larger open models on Microsoft's own OSWorld-Verified comparison.
What you could build with it
A small IT support team could run this model on a local Linux workstation to triage routine desktop chores: the agent reads the screen, and where a spreadsheet edit or file cleanup is faster in code, it switches to the terminal instead of clicking through menus.
Does it hold up?
Too early to judge for everyday use: weights and eval code are public, but independent reproductions are not out yet and local runtimes on Mac and Ollama could not yet load the model at the end of October.
Built with Mage-VL-Cowork
- Microsoft's Mage-VL-Cowork fits a desktop agent into 4B parametersi-SCOOP · Analysis of the hybrid GUI-plus-bash design, the Apache-2.0 release, and the caveats: vendor-reported scores, Linux-only evidence, and a research-only notice.
- Mage-VL-Cowork on a Mac: memory fits, runtime does not yetModelFit · Hands-on attempt showing the 2.7 GB quantized build fits a 16 GB Mac, but no stock llama.cpp, Ollama, or LM Studio build could load its vision projector as of 31 Oct.
- Microsoft releases Mage-VL-Cowork, a 4B computer-use modelXenoSpectrum · News coverage of the 66.0% OSWorld-Verified result and the teacher-harness evolution behind the training recipe.
Learn more
First spotted on hf: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73NVIDIA Isaac Lab 3.0 GA + Isaac OS open runtime + Agile ONE SN humanoidMajor Isaac platform expansion at GTC Berlin Oct 21: Isaac Lab 3.0 is generally available as the open-source robot…models · JEV 0.73NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MNVIDIA's Oct 21 collection drop for agentic retrieval: Nemotron-3.5-Embed in 8B BF16, 1B BF16 and 1B NVFP4…models · JEV 0.72Kimi K3.1 Pro: Moonshot's 2.8T open-weight flagship officially launchesMoonshot AI officially launched Kimi K3.1 Pro on Oct 23, a mixture-of-experts model with 2.8 trillion total / 104…models · JEV 0.72NVIDIA Cosmos 3.5 world models: now openly availableNVIDIA's Cosmos 3.5 family of open world foundation models for robots, autonomous vehicles and vision agents is now…models · JEV 0.71
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.