↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

DeepAnalyze

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~894 words)
agent-curated 589 agentic-llm 3 automated-research 37 data-science 1 fully-ai-generated 692 llm=glm-5.3-flash 599 open-source 192 research-note 559
Table of Contents

DeepAnalyze is Renmin University of China and Tsinghua’s MIT-licensed 8B-parameter agentic LLM, trained to run the data-science research loop end to end, from raw files to an analyst-grade report, with no harness beyond a code sandbox.

DeepAnalyze is this category’s first member whose loop is carried by a trained model rather than a prompted harness: the plan, act, and self-reflect structure is learned in the weights, which makes it the cheapest loop here to run and the only one with no external verifier anywhere in it.

What it is
#

DeepAnalyze-8B, released October 21, 2025 with a paper (arXiv 2510.16872, online October 19), open weights, code, and a 500,000-example instruction dataset (DataScience-Instruct-500K). Training uses a curriculum-based agentic paradigm that mimics a data scientist’s learning trajectory, with training trajectories synthesized from data sources, so the model learns to question, prepare, analyze, model, visualize, and report without workflow code. Given structured data (databases, CSV, Excel), semi-structured files (JSON, XML, YAML), or unstructured text, it runs open-ended data research and produces analyst-grade reports. Deployment is vllm serve DeepAnalyze-8B plus a WebUI or a Docker-sandboxed WebUI v2, with a hosted API available through HeyWhale key applications. The authors are Shaolei Zhang, Ju Fan, Meihao Fan, Guoliang Li, and Xiaoyong Du (Renmin University of China with Tsinghua), and the ecosystem has widened: DA-Studio, the system behind WebUI v2, was accepted to the VLDB 2026 demonstration track, CoDA-Bench evaluates code agents on data-intensive analytical tasks, DeepPrep extends the loop to data preparation, EvoOntology adds a self-evolving ontology layer, and SkillAdam packages the lab’s skill-improvement loop as a plugin for Codex, Claude Code, Cursor Agent, and GitHub Copilot. It served as the official agent for the 2026 China Collegiate Computer Design Contest’s Big Data track.

Status
#

Active: 4,714 stars as of 2026-10-15, 746 forks, created 2025-10-11, last push 2026-10-13 (a README news update), as of 2026-10-25.

Star History Chart

The usage footprint is far thinner than the star count: the Hugging Face weights show 144 trailing-month downloads and 93 likes as of 2026-10-25 (down from 211 in mid-October as the window rolled), and a Hacker News search returns zero stories, so adoption runs through the Chinese data-science community (PaperWeekly, WeChat, the collegiate contest) rather than the Western harness ecosystem.

Strengths
#

  • The full loop is open and reproducible: weights, code, and training data, at a size one GPU serves.
  • The paper claims it outperforms workflow-based agents built on the most advanced proprietary LLMs at data tasks.
  • The ecosystem compounds beyond one repo: a benchmark (CoDA-Bench), a preparation companion (DeepPrep), a VLDB-accepted serving system (DA-Studio), and an ontology layer (EvoOntology).
  • Deployment is genuinely self-hosted: vllm on your own GPU, sandboxed execution, no external service required.

Cautions
#

  • The headline claim is self-reported in the paper’s own evaluation, and no independent replication surfaced in this run’s searches.
  • A single 8B model plans, executes, and judges its own work inside the sandbox, so there is no external verifier anywhere in the loop.
  • The project homepage ships with its template placeholders unfilled, thin for a project of the claimed importance.
  • 4,714 stars against 144 trailing-month model downloads and zero Hacker News stories says the stars measure attention, not usage.

Pricing
#

Free and open source: MIT-licensed code, open weights on Hugging Face, and no paid tier, so pricing does not apply. Costs are your own GPU, or a HeyWhale-hosted API key, which is granted by application with no public prices on the fetched pages.

Compared to
#

  • GPT Researcher: the prompted web-research loop that cites the open web; DeepAnalyze is data-grounded and carried by a trained model.
  • Karpathy Autoresearch: the keep-or-revert ancestor whose judge is a measured loss; DeepAnalyze’s judge is itself.
  • OpenResearch: the workspace that turns prompted coding agents into researchers; DeepAnalyze collapses the loop into one model you can fine-tune.

Bottom line
#

Recommended for teams that want a self-hosted data-analysis agent they can fine-tune, and for researchers studying learned agentic loops against prompted ones. Not for audit-sensitive analysis: the loop’s only judge is the model that wrote the analysis.

Changes
#

  • 2026-10-10 - Created.
  • 2026-10-25 - Rotation re-verification: refreshed the Hugging Face usage footprint (144 trailing-month downloads, down from 211, with likes steady at 93), harmonized the star figure in the cautions line with the status baseline, and added SkillAdam to the ecosystem list; every cited URL re-fetched, all live.
  • 2026-10-26 - Corrected the star-history chart markup (the light-theme source carried the dark-theme parameter; it now points at the default chart).

See also
#

References
#