What is Magentic-UI (MagenticLite)?
Magentic-UI began in 2025 as a Microsoft Research web agent tuned for frontier models. In May 2026 version 0.2 replaced it with MagenticLite, which keeps the package name but is a different app: an orchestrator model, MagenticBrain (14 billion parameters, fine-tuned from Qwen 3 14B), plans the work and hands web steps to Fara1.5, a computer-use model in 4B, 9B and 27B sizes based on Qwen 3.5. The old frontier-model version lives on a separate magentic-ui-0.1 branch.
It runs on your own machine as a local web app, works across a browser and your file system, asks for approval before critical actions, and puts browser sessions in a lightweight VM sandbox called Quicksand. Microsoft calls all three components research releases, and the docs call MagenticLite a research prototype.
The code and both model weights are MIT-licensed. The latest PyPI release is 0.2.2, from 21 July 2026, and the last commit to the main branch was on 23 July 2026.
Is Magentic-UI (MagenticLite) worth it?
MagenticLite matters more as evidence than as a tool. Its point is that a 14B planner and a 9B browser model, both open under MIT, can handle web research, form filling and file chores without a frontier API. Microsoft's own figures support the direction: on Online-Mind2Web, a 300-task benchmark, it says Fara1.5 nearly doubles its predecessor Fara-7B and the 27B variant passes 90%. Those are self-reported results from the launch post, and the 9B model's score is shown only in a chart.
The practical cost is hosting. MagenticLite does not bundle models; it needs two OpenAI-compatible endpoints, one per role. The hosting guide's Hugging Face example suggests an L40S at about $1.80 an hour for Fara and an RTX PRO 6000 at about $2.75 an hour for MagenticBrain, about $4.55 an hour while both run, and scale-to-zero brings 30 to 90 second cold starts. Running them on your own GPUs means two vLLM servers, which the guide sets up on two single-GPU machines. Other models can be plugged in, but the docs warn the prompts are tuned for these two.
The safety design is the most transferable part. Browser sessions run inside the Quicksand VM, which needs KVM on Linux and WSL2, and the agent stops for approval before critical actions. The installation guide is blunt that the app exposes host file-system controls to anyone who can reach it, and that hosting it for several users is equivalent to handing out shell access.
The limitations page is unusually candid: summaries miss content, long conversations degrade, the planner can ignore a user's mid-task correction, file uploads through web forms do not work, and images cannot be task inputs. Add that the main branch has had no commit since 23 July 2026 and the version name is now carried by two different products, and this is a project to study, not to build a business process on.
Where it earns its place
An ML engineer at a bank that cannot send customer pages to a frontier API
Test whether open models can run internal web forms on-premises
Serve Fara1.5-9B and MagenticBrain with vLLM on two in-house GPUs and point MagenticLite at both. No page leaves the network, but the limitations page means a person should approve every submission.
A university lab studying agent safety
Compare how a small planner and browser model handle mid-task human corrections
Use the documented failure mode, where MagenticBrain can send Fara back to the original sub-task after a correction, as a test case. Hugging Face endpoints at about $4.55 an hour for both models keep a short study cheap.
An operations analyst at a 30-person firm with an existing GPU box
Gather supplier prices into a local spreadsheet each week
Run it in the default all mode so the browser model collects prices and the orchestrator writes the file. Expect to supervise; long sessions degrade and summaries can drop items.
Before you commit
- Setup
- uv venv --python=3.12, uv pip install "magentic_ui>=0.2.0" (pin it, or PyPI may serve 0.1.x), then magentic-ui --port 8081 and follow onboarding. Tested on Apple Silicon Macs and Windows 11 with WSL2; native Linux is untested, Windows ARM64 unsupported. Hugging Face endpoints take about 5 to 10 minutes to build.
- Integrations
- Any OpenAI-compatible /v1/chat/completions endpoint for each role; Azure OpenAI through config.yaml only. Three agent modes: both agents, orchestrator only for local tasks, or browser only.
- Data and privacy
- Runs locally; pages and files go only to the model endpoints you configure. The app exposes host file-system controls to whoever can reach it, so it should not be shared. Code and weights are MIT, with no revenue or use caps in the licence.
- Cost at scale
- On Hugging Face, about $1.80 an hour for Fara on an L40S and $2.75 for MagenticBrain on an RTX PRO 6000, roughly $800 a month if both run 8 hours a day for 22 working days. Self-hosting needs two GPU machines or one large enough for two vLLM servers.
- Limits
- About 20 open issues. Only one instance per port. No browser file uploads, no image inputs, weak long-document summarising, and planning that can override user steering. The 0.2 database does not migrate from 0.1.
What's good
- Code and both models under MIT
- Runs without frontier-model API calls
- Browser sandboxed in a local VM, with approval before critical actions
- Candid published limitations and transparency note
What's not
- Needs two separate GPU model endpoints
- Research prototype; no commits to main since July 2026
- Cannot upload files through web forms or take images as input
- 0.1 and 0.2 are different products under one package name
Best for
Researchers and platform engineers evaluating small open models for computer-use agents on hardware they control.
Magentic-UI (MagenticLite) alternatives
Magentic-UI (MagenticLite): common questions
- Is Magentic-UI (MagenticLite) worth it?
- MagenticLite is a worthwhile look at what small open models can do as a browser agent, but it is a research prototype that needs two GPU endpoints and has gone quiet since July. We rate it 3.4 out of 5.
- How much does Magentic-UI (MagenticLite) cost?
- Free under MIT, including the MagenticBrain and Fara1.5 weights. You host the two models yourself or on a GPU cloud.
- What are the downsides of Magentic-UI (MagenticLite)?
- Needs two separate GPU model endpoints; research prototype; no commits to main since July 2026; cannot upload files through web forms or take images as input; 0.1 and 0.2 are different products under one package name.
- Who is Magentic-UI (MagenticLite) best for?
- Researchers and platform engineers evaluating small open models for computer-use agents on hardware they control.
- How much does Magentic-UI (MagenticLite) cost for a team or in production?
- On Hugging Face, about $1.80 an hour for Fara on an L40S and $2.75 for MagenticBrain on an RTX PRO 6000, roughly $800 a month if both run 8 hours a day for 22 working days. Self-hosting needs two GPU machines or one large enough for two vLLM servers.
- Is Magentic-UI (MagenticLite) safe for company data?
- Runs locally; pages and files go only to the model endpoints you configure. The app exposes host file-system controls to whoever can reach it, so it should not be shared. Code and weights are MIT, with no revenue or use caps in the licence.
- What are the best Magentic-UI (MagenticLite) alternatives?
- The closest alternatives we have reviewed are Browser Use, Skyvern, BrowserOS and OpenHands.
Answers checked Oct 26, 2026.