Microsoft Releases Fara1.5: A Family of Browser Computer-Use Agents (4B/9B/27B) That Outperform OpenAI Operator and Gemini 2.5 Computer Use on Online-Mind2Web

Microsoft Research’s AI Frontiers lab launched Fara1.5. It is a household of computer-use agent (CUA) fashions for the browser. The launch ships three sizes: Fara1.5-4B, Fara1.5-9B, and Fara1.5-27B. The fashions are built-in with MagenticLite, Microsoft’s sandboxed browser interface for these brokers.

Computer-use brokers are pixel-to-action fashions that drive an actual browser. They learn screenshots and emit mouse and keyboard actions to finish duties. Recent agent merchandise like OpenAI’s Operator and Google’s Gemini 2.5 Computer Use sit on this class.

Fara1.5-27B scores 72% activity success on Online-Mind2Web. That benchmark covers 300 duties throughout 136 widespread websites. On the identical analysis, OpenAI’s Operator scores 58.3% and Gemini 2.5 Computer Use scores 57.3%. Yutori’s Navigator n1 reaches 64.7%, and Fara1.5-9B scores 63.4%. That almost doubles the predecessor Fara-7B, which scored 34.1% on the identical benchmark.

Architecture and agent loop

The fashions use Qwen3.5 base checkpoints of their 4B, 9B, and 27B variants. They function by an observe-think-act loop. At every step, the mannequin takes the prior dialog historical past and the three most up-to-date browser screenshots. It then emits ideas and a single subsequent motion.

The motion area contains normal mouse and keyboard inputs and web-specific actions like net search. It additionally exposes meta-actions for context administration. These embody memorizing info for later use and asking the consumer clarification questions. These meta-actions let the agent function over longer horizons and work collaboratively with customers.

Training combine

Training makes use of supervised fine-tuning on roughly two million samples. The combine is 60% net trajectories and 12.8% artificial environments. Form filling and consumer interactions account for 12.5%. Grounding contributes 8.8% and VQA 4.9%. Smaller slices cowl GUI drag, instruction following, and security. Loss is utilized solely to the three most up-to-date turns in every trajectory.

FaraGen1.5: the artificial information pipeline

FaraGen1.5 is the artificial pipeline that produced the coaching trajectories. It has three modular elements: environments, solvers, and verifiers.

Environments break up into two varieties. Open-internet duties run on reside web sites that don’t require logins. Gated-domain duties require authenticated classes or take irreversible actions, like sending an electronic mail.

For gated domains, the workforce constructed six artificial clones referred to as FaraEnvs. They cowl Mail, Calendar, Stream, ML, Stay, and Scheduler. Each clone has a practical frontend, a totally useful API, and a database with persona-based seed information.

These environments have been constructed utilizing GitHub Copilot CLI plus iterative human refinement. Because the workforce controls the complete stack, they know the proper end result for each activity. For duties that mutate the backend, an LLM decide compares database snapshots earlier than and after execution. Tasks that don’t change state are scored towards pre-computed reference solutions.

The solver agent makes use of OpenAI’s GPT-5.4 with customized instruments that mirror Fara1.5’s motion area. The solver scores 83% on Online-Mind2Web utilizing automated WebJudge. The earlier Fara-7B solver scored 67% on the identical analysis. A consumer simulator is invoked when the solver points an ask_user name or when it finishes a activity.

Three verifiers gate which trajectories enter coaching. Correctness makes use of LLM-generated rubrics for open-internet duties and privileged database judging for artificial ones. Efficiency penalizes redundant or pointless actions. User-interaction verification checks whether or not the agent paused at crucial factors.

Critical factors and security

Fara1.5 is skilled to cease and ask the consumer in three conditions. First: the duty requires private data the consumer has not supplied. Second: the duty description is ambiguous or lacking particulars wanted to behave. Third: an irreversible motion is about to be carried out with out prior approval.

Safety coaching makes use of public security datasets and inner duties aligned with Microsoft’s Responsible AI Policy. Inside MagenticLite, all agent actions are logged and auditable. The sandboxed browser additionally acts as a safety boundary between the agent and the consumer’s machine.

Other benchmarks

On WebVoyager, Fara1.5-27B scores 88.6%, the 9B reaches 86.6%, and the 4B hits 80.8%. The 9B additionally tops similar-sized friends like MolmoWeb 8B, GUI-Owl-1.5 8B, and Holo2 8B. All Fara1.5 analysis runs use Browserbase to stabilize classes and cut back session-level blocking. Numbers are averaged over three unbiased runs.

On WebTailBench v1.5, which targets long-tail net duties, Fara1.5-9B scores 64.5% course of success and 32.3% end result success. GPT-5.4 scores 79.6% course of and 57.4% end result on the identical benchmark.

Key Takeaways

Here are 5 one-line key takeaways:

• Microsoft Research launched Fara1.5, a household of browser computer-use brokers in 4B, 9B, and 27B sizes constructed on Qwen3.5.
• Fara1.5-27B scores 72% on Online-Mind2Web, beating OpenAI Operator (58.3%), Gemini 2.5 CU (57.3%), and Yutori Navigator n1 (64.7%).
• The FaraGen1.5 artificial information pipeline unlocks coaching on gated domains through six useful app clones (FaraEnvs) constructed with GitHub Copilot CLI.
• Fara1.5 pauses to ask the consumer at crucial factors: lacking information, ambiguous duties, or irreversible actions with out approval.

---

Check out the Technical details. Also, be at liberty to observe us on Twitter and don’t overlook to affix our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to accomplice with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so on.? Connect with us

The submit Microsoft Releases Fara1.5: A Family of Browser Computer-Use Agents (4B/9B/27B) That Outperform OpenAI Operator and Gemini 2.5 Computer Use on Online-Mind2Web appeared first on MarkTechPost.