Mage-VL-Cowork: Hardware Needed and How to Run It Free

Mage-VL-Cowork is a 4 billion-parameter model that can click and code within a single policy. It reaches 66.0 % success on the OSWorld-Verified benchmark, the highest score among 4 B models.

According to Microsoft, Mage-VL-Cowork is a multimodal hybrid computer-use agent built on the llava_onevision2 architecture. It combines a Mage-ViT visual encoder with the Qwen3-4B-Instruct-2507 language model, totaling 4 billion parameters and 9.48 GB in bfloat16 across two safetensors shards. The model accepts images and text and outputs text, supporting two action channels: a normalized GUI primitive set called computer_use and a persistent bash channel for shell and Python execution. It operates with a context window of 32,768 tokens, while the architecture can handle up to 262,144 tokens. Its asymmetric sliding-window memory keeps the three most recent screenshots and the four most recent full responses, plus folded action history, enabling long-horizon desktop tasks without exceeding the token budget.

What sets Mage-VL-Cowork apart is that the policy internally decides when to use the computer_use channel and when to switch to bash, removing the need for a separate multi-agent router. Microsoft evolved the teacher harness before large-scale trajectory collection, pruning redundant tools and pre-installing domain libraries such as python-pptx and pandas. The workflow-first engine, Mage-CUA-Forge, generates reusable skill-conditioned task workflows once per atomic skill, creating 293,806 distinct tasks across 6,915 skills while cutting recurring generation cost by 82.6 %. Training follows a five-stage specialize-then-unify curriculum, starting from an intermediate Mage-VL checkpoint and ending with online hybrid reinforcement learning that adds 10.16 percentage points over Hybrid SFT.

The model card reports that Mage-VL-Cowork achieves 66.0 % on the OSWorld-Verified benchmark under a 50-step budget, the highest score among 4 B models. On the OS task subset it reaches 87.5 %, and on multi-application workflows it scores 62.7 %. The headline figure is a two-run average across 361 held-out tasks, with runs of 65.33 % and 66.75 % respectively. Compared to other open-weight models at the same scale, such as ScaleCUA-Qwen3.5-9B (68.7 % overall) and ComputerRL-9B (48.0 % overall), Mage-VL-Cowork shows a clear advantage in both overall accuracy and specialized OS performance.

Mage-VL-Cowork’s main strength is its unified ability to perform visual GUI manipulation and structured terminal commands without an external router, which translates into high accuracy on OSWorld tasks. Online reinforcement learning reduces mean model calls from 21.57 to 18.77 per task and output tokens from 7,374 to 5,499, a 25.4 % token saving. Its 4 B size and 32,768-token context make it feasible for on-device use; the model runs on a single RTX 4060 with full weights in GPU memory. Limitations include the evaluation budget of 50 steps, meaning longer workflows have not been measured, and the lack of a free hosted API, users must run the weights locally. Performance may also vary on hardware that cannot hold the model in GPU memory.

The model is well suited for anyone who runs AI at home and wants a free, open-weight agent capable of both clicking and coding. Its Apache-2.0 license lets developers fine-tune or integrate it without restrictions. Because the full checkpoint fits on an RTX 4060, users with modest GPUs can experiment locally. Researchers exploring hybrid computer-use agents will benefit from the five-stage curriculum and the detailed teacher-harness evolution described in the model card. Anyone looking for a compact VLM that can handle desktop automation without paying for an API can download it from Hugging Face and start building custom workflows today.

Que faut-il pour le faire tourner chez vous ? En 4 bits, il prend environ 6 gigaoctets de mémoire. La plus petite machine personnelle qui le contient entièrement est RTX 4060. Sur RTX 3060 12 GB avec 32 gigaoctets de RAM, il écrit environ 65 tokens par seconde. Sur RTX 4090 avec 64 gigaoctets de RAM, il écrit environ 162 tokens par seconde.

Il n’existe pas encore d’API gratuite pour cela. Les poids sont ouverts : vous pouvez les télécharger depuis Hugging Face sous licence Apache 2.0. Arena ne l’a pas encore classé : nous ajouterons le score dès que les votes arriveront.

Compare Mage VL Cowork with every free model, and check what your own GPU can run, on FreeLLMArena.