November 26, 2025
Click, scroll, drama, repeat
Fara-7B by Microsoft: An agentic small language model designed for computer use
Microsoft’s tiny click-bot stirs hype, shade, and “can it play games” vibes
TLDR: Microsoft launched Fara-7B, a small on-device AI that literally clicks and types to handle web tasks and posts strong benchmark scores. The community is split: some love the local “AI intern,” others question synthetic training, Qwen roots, and whether anyone wants to outsource shopping—plus gamers want it flying rockets.
Microsoft dropped Fara-7B, a mini AI that uses your mouse and keyboard to do web chores—think clicking, scrolling, and filling forms like a speedy intern. It’s small (7B parameters), runs locally, and claims strong benchmark wins across shopping, travel, and multi-step tasks. The repo’s up at github.com/microsoft/fara, and yes, people are already trying it. But the crowd immediately split into camps: one cheering the on-device privacy and speed, another asking if this is just Qwen in a Microsoft hoodie. One blunt commenter flatly called it “fine tuned Qwen-7B,” while another poked at the corporate angle: why so much synthetic training data, and is Microsoft’s OpenAI deal limiting what they can do?
Meanwhile, the practical brigade showed up: “How much VRAM do I need?” asked a 12GB GPU owner, echoing the recurring confusion around local AI specs. The lifestyle angle sparked debate too—are people really outsourcing shopping to a bot? Some say yes for comparison shopping and form-filling; others see it as overkill for everyday tasks. And then came the memes: gamers begged for an AI pilot to play Kerbal Space Program, imagining Fara-7B yeeting rockets while it “learns” the UI. Love it or side-eye it, the vibe is clear: this click-bot has range—and baggage.
Key Points
- •Microsoft released Fara-7B, a 7B-parameter agentic model that controls mouse and keyboard to complete web tasks.
- •Fara-7B is trained on 145K synthetic trajectories using the Magentic-One framework and fine-tuned from Qwen2.5-VL-7B.
- •The model averages ~16 steps per task, compared to ~41 for comparable models, enabling efficient task completion.
- •Fara-7B shows strong benchmark performance (WebVoyager 73.5%, Online-M2W 34.1%, DeepShop 26.2%, WebTailBench 38.4%).
- •Microsoft introduced WebTailBench, a new benchmark with 609 tasks across 11 real-world categories, with detailed segment scores for Fara-7B.