MagenticLite's models are now fully open source. MagenticBrain and Fara 1.5, previously available on Microsoft Foundry, are live on Hugging Face with open weights. The app, the harness, and every model in the stack are now open.

Quick recap: MagenticLite is an open-source agentic application, the next generation of Magentic-UI. It works across the browser and local file system in a single workflow. It was built around a bet: small models plus a codesigned harness can match frontier-scale agentic systems.

The stack has two purpose-built models. Until today, both were only available on Foundry. Now you can pull the weights and run the full experience locally.

MagenticBrain is the orchestrator: a 14B model fine-tuned from Qwen 3 14B that plans, writes code, and delegates. It was trained end-to-end inside the MagenticLite harness, with the exact tool schemas it sees at inference. No gap between training and deployment.

Fara 1.5 is the computer-use specialist: 4B, 9B, and 27B models fine-tuned from Qwen 3.5. They read raw browser screenshots and emit clicks, typing, and scrolls. No DOM parsing, no accessibility trees.

The numbers, on Online-Mind2Web (300 tasks across 136 real sites): Fara1.5-27B: 72% OpenAI Operator: 58.3% Gemini 2.5 Computer Use: 57.3% The 9B lands at 63.4%, nearly double its predecessor Fara-7B.

And the 4B runs on-device at 57%. Pair it with MagenticBrain and MagenticLite's Quicksand sandbox, and your browser tasks, files, and data never leave your machine.

Everything in MagenticLite stays under your control: you see the agent's reasoning, you can take over at any moment, and critical actions pause for your approval. Now you can inspect the models behind it too.

Get started: Models: MagenticLite: Fara1.5: Fara1.5 blog: Technical report:

这下可以直接扒实现细节了

Open weights for the whole stack, harness included, is rare generosity.