仅27B参数，微软开源Fara 1.5超越OpenAI Operator：完整电脑操作Agent技术解析

by Mycelium Protocol

---

微软 Research AI Frontiers 团队在 2026 年 7 月发布了一个开源的电脑操作 Agent 模型家族 Fara 1.5，提供 4B、9B、27B 三种参数规模，全部以 MIT 授权释出，权重托管于 HuggingFace。

其中 Fara1.5-27B 在 Online-Mind2Web 基准检测拿下 72.3% 的高成功率，超越 OpenAI Operator 的 58.3% 与 Google Gemini 2.5 Computer Use 的 57.3%。9B 版本也达到 63.4%，在同等规模中刷新了 SOTA。

---

它能做什么

Fara 1.5 是原生电脑操作 Agent（Computer Use Agent，CUA）。给它一个任务描述，它直接在浏览器截图上观察当前状态，然后输出鼠标点击、键盘输入、搜索等操作——不需要 accessibility tree，不需要额外的页面解析模型，直接对坐标预测。

可以处理的任务类型包括：

• 搜索信息并整理结果
• 填写表单、管理账号设置
• 订票（机票、电影、餐厅）
• 跨站比价购物
• 查找职位、房源信息

循环逻辑是 observe → think → act：截图输入 + 对话历史 → 推理当前任务状态 → 输出下一步操作。

---

跑分数据

模型	规模	WebVoyager	Online-Mind2Web
OpenAI Operator	闭源	87.0	58.3
Gemini 2.5 Computer Use	闭源	—	57.3
GPT-5 SoM	闭源	90.6	57.7
Fara1.5-4B	4B	80.8	57.3
Fara1.5-9B	9B	86.6	63.4
Fara1.5-27B	27B	89.3	72.3

Fara1.5-9B 比上一代 Fara-7B 在 Online-Mind2Web 上提升了 29.3 个百分点。

---

技术核心：FaraGen1.5 数据管道

Fara 1.5 的核心不只是模型本身，而是支撑它的数据生成管道 FaraGen1.5，由三部分组成：

环境（Environments）

• 真实网站上的开放互联网任务
• 6 个合成 FaraEnv 环境（Mail、Calendar、Stream、ML、Stay、Scheduler）——这些是功能完整的 UI 克隆，用于模拟需要登录或会触发不可逆操作的场景，同时支持基于执行结果的精确验证

求解器（Solvers）

• 可接入多种模型，包括 GPT-5.4 等强 Frontier 模型
• 配备用户模拟器，支持多轮对话训练

验证器（Verifiers）

三层过滤：

1. 任务正确性（Universal Verifier LLM 裁判）
2. 效率（惩罚冗余操作）
3. 关键节点遵循（用户确认、缺失信息提示、不可逆操作暂停）

训练集最终约 200 万样本：~60% 真实网络轨迹、12.8% 合成环境、12.5% 表单填写、8.8% 基础定位、4.9% VQA、0.8% GUI 拖拽任务。

---

关键能力细节

直接坐标预测

Fara 1.5 不依赖 accessibility tree。它直接在截图像素上预测操作坐标，这意味着它能处理那些 DOM 结构混乱或根本没有 accessibility 支持的网页。

多轮用户交互

通过用户模拟器在多轮对话中训练，Fara 1.5 学会了：

• 主动询问缺失信息
• 标记模糊任务
• 在执行不可逆操作前暂停并请求用户确认

这解决了自动化 Agent 最容易出问题的场景——“你让它订机票，它直接付款了”。

可审计性

与 Magentic-UI 集成后，所有操作都有日志可追溯，并在关键节点弹出用户确认界面。

---

快速开始

# 1. 克隆并安装
git clone https://github.com/microsoft/fara.git
cd fara
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
playwright install

# 2. 从 Microsoft Foundry 部署 Fara1.5-9B，创建配置文件
cat > azure_foundry_config.json << 'EOF'
{
    "model": "Fara1.5-9B",
    "base_url": "https://your-endpoint.inference.ml.azure.com/",
    "api_key": "YOUR_API_KEY_HERE"
}
EOF

# 3. 运行任务
fara-cli --task "What's the weather in New York now?" \
         --endpoint_config azure_foundry_config.json

或者在 Magentic-UI 中以可视化方式使用，支持沙箱浏览器环境 + 操作日志界面。

---

两个配套基准

WebTailBench：609 个任务，覆盖 11 类真实网络任务，重点测试现有基准忽视的长尾场景（购物对比、跨站任务、多步组合）。数据集在 HuggingFace：microsoft/WebTailBench。

CUAVerifierBench：专门评估”给 Agent 打分的裁判”质量。包含 Fara Agent 轨迹 + 人工标注判断 + Universal Verifier 输出，用于开发和比较 Agent 评估方法。

---

为什么值得关注

Fara 1.5 有几个值得注意的点：

27B 就能超越闭源大厂。这不是微调一个聊天模型，而是专门训练的 CUA 模型，规模 27B，成本远低于调用 OpenAI Operator。

方法论开放。FaraGen1.5 管道（环境 + 求解器 + 验证器）的设计细节公开发表，任何团队都可以基于这个框架构建自己的垂直领域 Agent 训练数据。

坐标直接预测 vs. accessibility tree。这个选择意味着 Fara 1.5 能处理 accessibility tree 不完整或不可用的场景，覆盖范围更广。

仓库：github.com/microsoft/fara · HuggingFace：microsoft/Mage-VL · 论文：arxiv 2606.20785

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Microsoft Fara 1.5: A 27B Open-Source Agent That Outperforms OpenAI Operator

by Mycelium Protocol

Microsoft Research AI Frontiers released Fara 1.5 in July 2026 — a family of native computer use agent (CUA) models at three scales: 4B, 9B, and 27B. All three are MIT-licensed and available on HuggingFace.

Fara1.5-27B achieves 72.3% on Online-Mind2Web, outperforming OpenAI Operator (58.3%) and Gemini 2.5 Computer Use (57.3%). The 9B version reaches 63.4% — a new state of the art for its size class, and a +29.3 point improvement over the previous Fara-7B.

What It Does

Fara 1.5 is a native Computer Use Agent. Given a task description, it observes the current browser state through a screenshot, reasons about what to do next, and outputs mouse clicks, keyboard inputs, or web searches — with no accessibility tree, no separate parsing model, direct coordinate prediction on the screenshot.

Supported task types: information search and summarization, form filling, account management, booking (flights, restaurants, movie tickets), cross-retailer price comparison, job and real estate searches.

The loop: observe (screenshot + conversation history) → think → act.

Performance

Model	Size	WebVoyager	Online-Mind2Web
OpenAI Operator	closed	87.0	58.3
Gemini 2.5 CU	closed	—	57.3
GPT-5 SoM	closed	90.6	57.7
Fara1.5-4B	4B	80.8	57.3
Fara1.5-9B	9B	86.6	63.4
Fara1.5-27B	27B	89.3	72.3

FaraGen1.5: The Data Pipeline

The core innovation behind Fara 1.5 is FaraGen1.5, a scalable training data pipeline with three modular components:

Environments — open-internet tasks on live websites plus six synthetic FaraEnvs (Mail, Calendar, Stream, ML, Stay, Scheduler). These are functional UI clones that simulate authentication-gated or irreversible-action domains while enabling ground-truth execution-based verification.

Solvers — a solver harness powered by strong frontier models (including GPT-5.4) paired with a user simulator for multi-turn rollouts.

Verifiers — three complementary filters: task correctness (Universal Verifier LLM judge), efficiency (penalizing redundant actions), and critical-point adherence (flagging missing info, ambiguous tasks, unapproved irreversible actions).

Training mix: ~2M samples — 60% web trajectories, 12.8% synthetic environments, 12.5% form filling, 8.8% grounding, 4.9% VQA, 0.8% GUI drag tasks.

Key Technical Choices

Direct coordinate prediction on screenshots. No accessibility tree dependency. This means Fara 1.5 handles pages with incomplete or absent accessibility support — a significant portion of the real web.

Multi-turn user interaction via user simulator. Trained to ask for missing information, flag ambiguous tasks, and pause before irreversible actions. This is the critical safety behavior that distinguishes a deployable agent from a dangerous one.

Auditable action logging via Magentic-UI. A sandboxed browser environment with user confirmation prompts at critical points and full action logs. For users who want automation without blind trust.

Quick Start

git clone https://github.com/microsoft/fara.git && cd fara
python3 -m venv .venv && source .venv/bin/activate
pip install -e . && playwright install
# Deploy Fara1.5-9B from Microsoft Foundry, create config, then:
fara-cli --task "Find flights from SF to NYC next Friday" \
         --endpoint_config azure_foundry_config.json

Or use Magentic-UI for a visual interface with action logging.

Two Companion Benchmarks

WebTailBench (609 tasks, 11 real-world task types): focuses on long-tail scenarios underrepresented in existing benchmarks — shopping lists, comparison shopping, compositional cross-site tasks. Dataset: microsoft/WebTailBench on HuggingFace.

CUAVerifierBench: evaluates the judges that score CUA agents, not the agents themselves. Pairs Fara trajectories with human verdicts and Universal Verifier outputs to enable systematic development of better evaluation methods.

Why This Matters

A 27B open-source model outperforming closed proprietary systems on a standardized web-automation benchmark is a meaningful signal. The FaraGen1.5 pipeline design — environments + solvers + verifiers — is fully described in the paper and reproducible: teams can adapt it for vertical domains without starting from scratch.

The choice to avoid accessibility trees broadens applicability. The built-in user confirmation behaviors address the safety concerns that make autonomous agents risky in practice.

Repository: github.com/microsoft/fara · HuggingFace: aka.ms/fara1.5-hf · Paper: arxiv 2606.20785

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.