Microsoft launches Fara-7B, an agentic Qwen model that solves tasks in ~16 steps
When Microsoft announced Fara-7B this week, I was surprised to see it billed as an “agentic” model that can actually run computer-based tasks. Under the hood it runs on Qwen2.5-VL-7B, a visual-language core that the team has fine-tuned with a supervised approach. To teach it, they apparently cranked out about 145 000 synthetic trajectories using their Magentic-One framework and then fed those examples straight into the model.
It feels like a clear sign that Microsoft wants to bundle more capable, end-to-end agents into enterprise workflows, where speed and reliability matter just as much as raw power. The tech looks solid, but I’m still not sure how it will compare with the dozens of similar agents already out there. Microsoft claims the model hits impressive numbers on speed and accuracy, yet the real test will be how it performs in everyday business settings.
Here’s what the company says about its performance:
Microsoft says the model finishes tasks in about 16 steps on average, which is far fewer than many comparable systems. The model is trained on 145,000 synthetic trajectories generated through the Magentic-One framework and is built on Qwen2.5-VL-7B with supervised fine-tuning. The company positions Fara-7B as an everyday computer-use agent that can search, summarise, fill forms, manage accounts, book tickets, shop online, compare prices and find jobs or real estate listings.
Microsoft is also releasing WebTailBench, a new test set with 609 real-world tasks across 11 categories. Fara-7B leads all computer-use models across every segment, including shopping, flights, hotels, restaurants and multi-step comparison tasks. The company offers two ways to run the model.
Azure Foundry hosting lets users deploy Fara-7B without downloading weights or using their own GPUs. Advanced users can self-host through VLLM on GPU hardware. The evaluation stack relies on Playwright and an abstract agent interface that can plug in any model.
Microsoft warns that Fara-7B is an experimental release and should be run in sandboxed settings without sensitive data. Earlier this year, Microsoft launched Phi-4-multimodal and Phi-4-mini, the latest additions to its Phi family of small language models (SLMs).
Microsoft claims its Fara-7B, a 7-billion-parameter model, can hold its own against much larger agents. In their demo the system finishes a live web task in about sixteen steps, which is noticeably fewer than many competitors. It “sees” the page as an image, then predicts where to click, type or scroll, sidestepping accessibility trees and extra parsing.
The backbone is Qwen2.5-VL-7B, fine-tuned on roughly 145 000 synthetic trajectories generated by the Magentic-One framework, and it runs locally, so latency is low and privacy is better, at least according to Microsoft. The bigger question is whether the numbers actually hold up. No public benchmark or third-party test has been released, so we can’t say how it behaves on the messier tasks users throw at it.
Relying on synthetic data also makes me wonder how well it will cope with truly unexpected inputs. Still, the work shows a modest-size model can be built for direct computer interaction without a massive backend. Whether developers will adopt it will probably hinge on more open evaluation and real-world feedback.
Related Reading
Browse all Microsoft news →
Business & Startups Oracle, NVIDIA deepen tie-up to speed sovereign AI and government digital shift
Business & Startups Skilling programs lag AI; students must prioritize aspiration and depth
Business & Startups EPAM and Cursor partner to scale AI coding for global enterprise customers
Related Topic LG's recent webOS update adds Microsoft Copilot app, now removable
AI Tools & Apps
Related Topic Microsoft partners with PayPal to add AI shopping and payments in Copilot
AI Tools & Apps
Common Questions Answered
What is the base model for Microsoft’s Fara-7B and how was it adapted?
Fara-7B is built on the Qwen2.5-VL-7B visual‑language backbone. Microsoft fine‑tuned this model with supervised learning using about 145,000 synthetic trajectories generated by the Magentic‑One framework.
How many steps does Fara-7B typically need to complete a task, and why is this notable?
Microsoft reports that Fara-7B finishes tasks in roughly 16 steps on average. This is notable because it is significantly fewer steps than many comparable agentic systems, indicating higher efficiency in end‑to‑end workflows.
What types of computer‑based tasks is Fara-7B designed to handle?
Fara-7B is positioned as an everyday computer‑use agent capable of searching the web, summarising content, filling forms, managing accounts, booking tickets, shopping online, comparing prices, and locating jobs or real‑estate listings.
How does Fara-7B interact with web pages differently from traditional agents?
Instead of relying on accessibility trees or extra parsing layers, Fara-7B reads pages visually and interacts by predicting coordinates for clicks, typing, and scrolling. This visual approach allows the model to operate directly on rendered page content.
What advantage does running Fara-7B locally provide for enterprise users?
Running locally reduces latency compared to cloud‑only solutions and gives enterprises more control over data privacy. The 7‑billion‑parameter size also makes it more resource‑efficient while still delivering agentic capabilities.