最后更新：10/24/2026最后核验：2026-10-24
广告
<p>Cloudy 将长时间运行的机器学习实验交给云端研究智能体：用户描述研究目标并附上代码或数据，系统规划实验、编写和执行代码、读取测量结果，再决定下一轮尝试。它可以按步骤切换 CPU 与最多八块 NVIDIA GPU，保留实验代码、日志、图表、结果和检查点，并在同一对话中恢复研究。用户可连接 Claude Code 或 Codex 订阅承担模型用量，也可使用 Cloudy 预付积分；计算、搜索和存储仍从团队积分扣费。预付积分从 5 美元起，无订阅费，平台费和税费在结账时展示。研究时长上限不是金额预算，暂停会保存文件和检查点，但不会让运行中的进程继续存活。本文依据官方公开资料核验，未注册账号或运行付费实验。</p><ul><li>从目标和用户代码出发，自主规划并迭代机器学习实验</li><li>按需使用云端 CPU 与最多八块 NVIDIA GPU</li><li>保存文件、结果和检查点，可在同一对话中恢复研究</li><li>交付代码、日志、测量数据、图表与有证据支撑的研究报告</li></ul>
详情
平台
网页版macOSLinuxWindows
功能特性
- Plans and iterates on ML experiments from a goal and user-provided code or data
- Runs code on cloud CPUs and up to eight NVIDIA GPUs, switching hardware as needed
- Saves files, results, and checkpoints for resuming research in the same conversation
- Returns code, logs, measurements, plots, and an evidence-backed research report
支持语言
en
适合人群
ML engineers running training, fine-tuning, or optimization experiments without their own GPUsStartup research teams that need long-running iterative experiments with saved resultsDevelopers delegating measurable ML research goals and reviewing the resulting code and evidence
已知限制
- Uses prepaid team credits starting at $5; compute, search, and storage remain chargeable even with a connected Claude or ChatGPT plan
- Research duration is a time limit, not a dollar spending cap; platform fees and taxes are shown at checkout
- Paused research preserves files and checkpoints but running processes do not stay alive while paused
- Requires an account; the optional CLI needs Python 3.12 or later and Windows users must use WSL
替代工具
Claude Code — Terminal-first coding agent when cloud GPU experiment orchestration is not needed
OpenAI Codex — Coding agent for repository tasks rather than an autonomous GPU research workflow
Julius AI — Conversational data analysis rather than long-running ML experiment optimization
