MagenticLite and Fara1.5: Microsoft Revolutionizes Small AI Models

MagenticLite: A New Era for Agentic AI

MagenticLite is an agentic application that marks a significant advancement in the field of small artificial intelligence models. Developed by Microsoft Research AI Frontiers, this application is designed to operate seamlessly both in web browsers and on local file systems. It represents the next generation of Magentic-UI, integrating a specially optimized environment for small models, enabling complex tasks to be performed with increased efficiency.

The main innovation of MagenticLite lies in its seamless integration with two specific models: MagenticBrain and Fara1.5. MagenticBrain is designed for orchestration and reasoning, while Fara1.5 specializes in browser-based computing tasks. Together, these models demonstrate how co-designed tools and an optimized execution environment can push the boundaries of agentic performance, even with smaller-sized models.

Today, Microsoft Research AI Frontiers is releasing MagenticLite, an experimental agentic application designed for small models. As the next generation of Magentic-UI, it operates across both the browser and the local file system in a unique workflow.

MagenticLite is powered by two specially designed models: MagenticBrain, for reasoning, delegation, and terminal use, and Fara1.5, a family of models for browser-based tasks. The three components have been designed to work together as a single system. The result is an agent that operates efficiently, retains data on the user's machine, and supports a wide range of agentic tasks. It also points towards a broader goal: agents capable of functioning directly on users' hardware.

The project is based on a key research bet: that agentic capability depends on the orchestration of tools and action rather than knowledge alone. This understanding allows for the use of smaller models while enabling a wide range of agentic tasks at a fraction of the cost.

MagenticLite also reflects our end-to-end approach to agentic AI — from generating training data and designing models to orchestration, interaction design, and human oversight throughout the experience.

Included in this Release

MagenticLite

The next generation of Magentic-UI, our experimental agentic experience, is powered by a rebuilt agent environment for small models, featuring an updated user interface based on community feedback. It operates across users' browsers and local file systems in a unique workflow.

MagenticBrain

MagenticBrain is the planner, coder, and delegator of MagenticLite all in one. It transforms vague requests into concrete plans, selects the right tool or sub-agent for each step, writes code when necessary, and gets back on track if something fails during the task.

Fara1.5

The next generation of our family of computing models, Fara1.5, is available in three sizes, with a flagship model of 9 billion parameters recommended for most use cases. Fara1.5 sets new state-of-the-art (SOTA) results among small computing models and nearly doubles the performance of Fara-7B on web navigation, with more precise handling of forms, sites requiring credentials, and long-duration tasks.

Each component is useful on its own, but they work best together. The co-design of the application, models, and environment enables capable and reliable agentic performance at this scale.

Our Research Approach: Doing More with Less

We started with a simple question: what does it take for a small model to be truly good at agentic tasks? The answer encompassed the entire lifecycle — data generation, training objectives, model design, and orchestration needed to be redesigned together rather than in isolation.

We identified requirements from real-world use cases such as form filling, browser searching, and local file management, and built an evaluation dataset around them. Standard benchmarks capture part of the picture but do not always directly measure real-world utility. Scenario-based assessments complemented these benchmarks and became a key signal for the iterative improvement of models and the environment.

For the user experience, we retained key elements from Magentic-UI, including visibility into the agent's reasoning and actions, the ability for users to take direct control, and explicit approval at critical points. Based on recent studies, we also made MagenticLite easier to learn and use with updated views for the browser and chat, designed to facilitate understanding of the agent's actions and intervention when necessary.

System Components

Fara1.5: A Computing Model That Outperforms Its Category

Fara1.5 is the next generation of our family of computing models, available in three sizes, with a flagship model of 9 billion parameters recommended for most use cases. Fara1.5 achieves new SOTA results among small computing models and nearly doubles the performance of Fara-7B on web navigation, with better handling of forms, sites requiring credentials, and long-duration tasks.

Last November, we released Fara-7B, a small agentic model designed to perform tasks in a web browser. It was trained using a new synthetic data generation engine that enabled top-tier performance. Fara1.5 is the next step in this bet: a family of three models (4B, 9B, 27B) based on Qwen 3.5, designed to fill the gaps we observed in the previous version.

State-of-the-Art Results

On the popular benchmark Online-Mind2Web, which contains 300 tasks across widely used web domains, Fara1.5 sets new SOTA results for models in its category. Fara1.5 outperforms all similarly sized models and nearly doubles the performance of Fara-7B. The larger variant Fara1.5-27B achieves over 90% performance on the same benchmark.

Enhanced User Experience

In addition to improvements on benchmarks, we have enhanced the user experience of Fara1.5. Users should observe stronger performance on everyday tasks such as form filling, managing logins for credential-required sites, and booking appointments. These improvements are powered by the next evolution of our FaraGen data generation pipeline. In addition to training on live websites, we also trained the model in highly realistic synthetic environments designed to simulate scenarios like logins and irreversible actions.

A Native Action Space Optimized for Long-Duration Tasks

Beyond clicks and keyboard actions, Fara1.5 has built-in tools to store key information in its context across hundreds of steps and ask the user for their permission or preferences when necessary, helping it remain consistent on tasks that span many minutes of real work.

Recalibrated Critical Points

Fara-7B was trained to detect critical points for activities such as transactions, login flows, or irreversible submissions and to signal them. In Fara1.5, we refined our design around critical points based on our learnings from real-world usage, ensuring that security triggers occur when they should but do not block useful tasks, like form filling.

MagenticBrain: The Orchestration Model

MagenticBrain is a 14 billion parameter orchestration model — planner, coder, and delegator all in one. Fine-tuned from Qwen 3 14B, MagenticBrain was trained end-to-end within the MagenticLite environment with the same tool schemas and execution environment it will encounter at inference time.

In many agentic systems, orchestration (planning and coordination) is the most reasoning-intensive component, so teams have historically relied on their most capable models for this role. Our bet is that small models can take on this role without sacrificing their capability. Two design choices make this possible.

The first is combining multi-step tool call trajectories — where the model learns to choose the right tool and call it correctly — with coding and terminal trajectories — where the right answer is sometimes five lines of Python, not a tool call. This is associated with a close coupling between the tool format used during training and inference.

The second is the delegation of computer usage agent (CUA). A key part of the orchestrator's job is knowing when not to act itself and to delegate a task to Fara1.5. Our data pipeline includes explicit delegation trajectories: sequences where the orchestrator recognizes a browser or user interface (UI) task, issues a structured handoff to the CUA model, waits for the result, and resumes the task. The result is an orchestration model that reasons, codes, calls tools, and delegates seamlessly within a single 14 billion parameter space. We are releasing MagenticBrain, designed to be used with MagenticLite.

The Environment: Designed for Small Models

The environment combines orchestration and browser usage models into a single workflow. Three design choices are most important:

• Step-by-Step Planning: The environment plans incrementally, keeping the system flexible and allowing for smoother course corrections and recoveries throughout long-duration tasks.

• Active Context Management: Small models have smaller effective context windows and degrade more quickly as context grows. The environment actively manages what each model receives at each step, keeping prompts targeted, only surfacing necessary information, condensing previous interactions into concise summaries, and offloading the rest, so that the orchestrator and Fara1.5 remain efficient on long tasks.

• Delegation via Sub-Agents: Rather than relying on a single small model for each task, the orchestrator acts.