Renmin University's DeepAnalyze Beats GPT-4 at Data Science With 8B Parameters

A team from Renmin University of China and Tsinghua has released DeepAnalyze, an 8B parameter open-source model that positions itself as the first agentic LLM built specifically for autonomous data science. Instead of chaining a general-purpose LLM through a hand-written workflow, the model is trained to decide, on its own, when to inspect data, when to write code, when to execute it, and when to write up findings.

DeepAnalyze-8B ships with weights, training code, a 500K trajectory dataset, and multiple frontends (WebUI, JupyterUI, CLI, OpenAI-style API) all under MIT license. Early community reactions highlight the fact that an 8B model is beating agents built on frontier proprietary LLMs on multiple data science benchmarks.

What it actually does

The model can autonomously complete a wide range of data-centric tasks without human intervention, covering the entire pipeline (data preparation, analysis, modeling, visualization, and report generation) as well as open-ended deep research over structured data (databases, CSV, Excel), semi-structured data (JSON, XML, YAML), and unstructured data (TXT, Markdown), producing analyst-grade research reports.

You point it at a folder of files with a task instruction, and it iterates until it produces a written report with charts. A minimal call looks like this:

from deepanalyze import DeepAnalyzeVLLM
deepanalyze = DeepAnalyzeVLLM("path/to/deepanalyze-8b/")
answer = deepanalyze.generate(
    prompt="Generate a data science report.\n# Data\nFile 1: person.csv ...",
    workspace="/path/to/data/"
)
print(answer["reasoning"])

The five-action loop

Where a typical data agent uses a fixed plan-code-execute template, DeepAnalyze is trained to emit one of five action tags at each step. It uses five actions, Analyze, Understand, Code, Execute, and Answer, so the model can switch between thinking and doing. The JupyterUI demo makes this concrete: