DeepAnalyze is an open-source agentic large language model that autonomously carries out data science work — data preparation, analysis, modeling, visualization, and report generation — for data analysts, researchers, and engineering teams that want to run those pipelines on their own infrastructure instead of through a hosted assistant.

What it is

DeepAnalyze is presented as the first agentic LLM for autonomous data science. It is developed by authors Shaolei Zhang, Ju Fan, Meihao Fan, Guoliang Li, and Xiaoyong Du, affiliated with Renmin University of China and Tsinghua University, and it is documented in the paper arXiv 2510.16872. The released model is DeepAnalyze-8B, published on Hugging Face under RUC-DataLab. The project is written in Python and licensed under MIT, with 4,636 stars, 735 forks, and 27 open issues in its GitHub repository. Model, code, training data, and demo are all open-sourced, and the project's stated goal is that a user can deploy or extend their own data analysis assistant.

The concrete problem it addresses is the manual stitching that normally surrounds analytical work: pulling data out of a database or a CSV file, cleaning it, modeling it, charting it, and then writing the findings up by hand. DeepAnalyze claims to run that entire sequence autonomously, without human intervention between steps, and to close the loop by producing an analyst-grade research report rather than a loose collection of intermediate artifacts. Because the model weights, the repository, the dataset, and the demo are all public, the project also replaces the alternative of relying on a closed hosted assistant for the same work, giving teams the option to inspect, self-host, or fine-tune the component that performs the analysis.

Key capabilities

• Runs an entire data science pipeline autonomously: data preparation, analysis, modeling, visualization, and report generation, without human intervention.
• Conducts open-ended data research across structured sources (databases, CSV, Excel), semi-structured sources (JSON, XML, YAML), and unstructured sources (TXT, Markdown).
• Emits analyst-grade research reports as the final output of that research, rather than only intermediate code or charts.
• Ships as DeepAnalyze-8B on Hugging Face, with a hosted demo on the same model page.
• Releases training data as DataScience-Instruct-500K under RUC-DataLab, alongside the model and code.
• Published under the MIT licence in Python, with a public repository and a project homepage at https://ruc-deepanalyze.github.io.
• Adds SkillAdam, released in September 2026, a plugin that generates tasks, evaluates a current skill, proposes improvements, validates the revision, and updates the skill when changes pass evaluation.

Who uses it and how

• Analyst and research teams that want an assistant they can self-host and extend, since weights, code, and training data are all available for local deployment or further training.
• Research groups working across mixed formats, where a single request may touch a database, a CSV or Excel sheet, a JSON or YAML configuration, and Markdown or plain text notes before a report is produced.
• Agent builders who already work inside Codex, Claude Code, Cursor Agent, or GitHub Copilot and want to improve an agent skill automatically through SkillAdam, supplying a SKILL.md and a description of the desired improvement.
• Report-writing workflows that expect a finished analyst-grade document as the deliverable rather than a notebook of intermediate steps.
• Projects with an existing audience and community footprint: the repository has accumulated 4,636 stars and 735 forks.

Getting started

The README excerpt provides no package name, Docker image, or compose file; the documented entry points are the Hugging Face model RUC-DataLab/DeepAnalyze-8B, its demo, the GitHub code repository, and the DataScience-Instruct-500K dataset, with the homepage at https://ruc-deepanalyze.github.io collecting cases.

How it compares

The facts provided name no paid products that DeepAnalyze replaces, and they name no directly comparable open-source alternatives. On the evidence given here, it stands alone in this registry as an agentic model for autonomous data science.

When to use it — and when not to

Choose it when the goal is an autonomous analytical pipeline over heterogeneous files and databases and the team is willing to run or extend an 8B model itself. Avoid it if a turnkey hosted service is expected: the README excerpt shown here carries no installation, packaging, or deployment commands, so the practical starting point is the Hugging Face model card and repository rather than a one-line install. Note also that the excerpt is truncated, and apart from the 2026.09 SkillAdam release there is no release cadence stated beyond the last repository push.