Open your data locally
Start with CSV, Excel, experiment files, or a SQL database already available from your computer. There is no need to paste hundreds of rows into a chat window.
sales.xlsx · customers.csv · experiment-results.csv
Local-first AI data analysis
Data Analysis with AI + Python Using Local LLMs
Analyze your data with AI without giving up control of your files.
Ask questions in plain English, run AI-generated Python locally, inspect every step, and keep the complete workflow in a reproducible notebook. MLJAR Studio currently supports local models through Ollama.
Local Python execution Visible generated code Notebook output
Runs locally
Your files
CSV · Excel · SQL
MLJAR Studio
Desktop workspace
Local LLM
Understands and writes code
Python
Calculates the real result
Reproducible notebook
Question + code + output + chart
The basic idea
What is local AI data analysis?
A language model running on infrastructure you control works alongside Python, pandas, visualization libraries, and statistical packages.
Instead of asking an LLM to calculate everything in a text response, the model helps decide what analysis to perform and generates the Python code needed to do it. Python then runs against your actual data.
The distinction matters
The LLM handles reasoning, instructions, and code generation. Python performs transformations, statistics, visualizations, and machine learning.
In MLJAR Studio, generated Python is visible and every executed step can be preserved in the notebook, making the analysis inspectable and reproducible.
Your data
Files or SQL
Local LLM
Plans and writes Python
Notebook
Code, output, explanation
Verifiable answers
Why use Python together with a local LLM?
An LLM can translate an analytical question into a plan and code. Python makes sure the answer is calculated from the real data.
“Which customers increased their spending the most compared with last year?”
You usually do not want the language model to guess the answer. You want it to create code that reads the data and calculates the result. The AI makes Python easier to use; Python makes the analysis verifiable and repeatable.
Load CSV and Excel files
Query databases
Clean inconsistent data
Merge datasets
Calculate statistics
Group and aggregate records
Create charts
Detect outliers
Train machine learning models
Export data and reports
From question to notebook
How local LLM data analysis works
The language model and Python have separate jobs, and every stage stays available for inspection.
Ask a question in plain English
Describe the result you need: compare groups, highlight a decline, merge files, clean records, or create a visualization.
“Show revenue by month and highlight the three largest declines.”
The local LLM generates Python
The model interprets the request, plans the analysis, and writes Python using pandas and the data-science libraries suited to the task.
Readable code remains available for review and editing.
Python runs the calculation
Python executes against the real dataset and produces a table, statistic, chart, cleaned file, or machine learning model.
The result comes from computation, not a language-model guess.
Keep the complete workflow
Questions, generated code, outputs, charts, and follow-up analysis stay together in a notebook that can be reopened and rerun.
Inspectable · editable · reproducible
generated-analysis.py
monthly = (
df.groupby("month", as_index=False)["revenue"]
.sum()
.sort_values("month")
)
monthly["change"] = monthly["revenue"].pct_change()
monthly.nsmallest(3, "change")Python groups the actual data, calculates month-over-month change, and returns the three largest declines.
Use cases
What can you analyze with a local LLM and Python?
Use the same conversational workflow for everyday data preparation, rigorous analysis, and predictive modeling.
Excel and CSV files
Clean spreadsheets, merge files, find duplicates, summarize columns, calculate KPIs, create charts, and detect unusual values.
Business and database data
Combine natural-language questions with SQL and Python, then continue the analysis after the selected data is loaded.
Statistical analysis
Use visible Python for t-tests, ANOVA, correlations, regression, confidence intervals, and outlier detection.
Data visualization
Ask for trends, distributions, comparisons, or relationships and let Python create the visualization from your data.
Machine learning
Prepare features, train models, evaluate results, and continue experimenting without moving the data to another tool.
Privacy and control
Why run the LLM locally?
The biggest difference between a local and cloud LLM is where inference happens.
With a local model, inference can run on your computer or infrastructure you control. That gives you more control over model choice and can be useful when data cannot be uploaded to a third-party AI service.
Confidential company data
Unpublished research
Financial information
Customer records
Healthcare datasets
Proprietary databases
Know the tradeoffs
Local control uses local resources
Cloud models provide capable inference without requiring powerful local hardware. A local model needs enough RAM or GPU memory for the model you want to run, plus memory for Python and your dataset.
Your hardware defines the practical model size
Smaller models are easier to run; larger models often improve analysis quality but demand more memory.
Two separate choices
Local LLMs are not the same as local LLM providers
First choose the neural network that will generate responses. Then choose the software that loads and serves it.
1. Choose a model
The model is the neural network that understands your request and generates the response. Families such as Qwen and Gemma are examples used in current MLJAR Ollama documentation.
Python coding abilityReasoning and tool useContext sizeSpeedMemory requirementsModel license
2. Choose a model runtime
The runtime loads the model and makes it available to applications. These tools solve model serving; MLJAR Studio turns the served model into a practical Python data-analysis workflow.
Ollama
SupportedA straightforward way to download and run models locally. It is the local provider currently documented and supported by MLJAR Studio.
LM Studio
Runs models locally and can expose local REST and OpenAI-compatible API endpoints.
Jan
Includes a local OpenAI-compatible API server powered by llama.cpp.
vLLM
Provides an OpenAI-compatible server for serving models from GPU infrastructure.
Choosing a model
Which local LLM is best for data analysis?
There is no single best model for every computer and every analytical task.
A small model may run comfortably on a laptop but produce weaker code. A larger model may perform better but require considerably more memory. Model selection should be based on real data-analysis tasks, not only generic chatbot benchmarks.
- 1Understand the analytical request
- 2Generate correct Python
- 3Work with pandas and data-science libraries
- 4Interpret outputs from previous steps
- 5Recover when generated code fails
- 6Continue a multi-step analysis
- 7Explain the final result clearly
MLJAR LLM benchmark
We benchmark LLMs on real Python data analysis
Every model receives the same multi-step analytical prompts. The complete interaction is saved as a notebook, so you can inspect prompts, generated Python, outputs, errors, charts, final answers, and scores.
Exploratory data analysis
Statistics
Machine learning
Time series
Finance
NLP
Local resources
What hardware do you need?
The practical answer depends mainly on the model, but the language model is only part of your memory budget.
RAM
System memory used by the model, Python environment, and your dataset.
VRAM
GPU memory available for acceleration when your machine has a compatible GPU.
Model size
Larger models generally need more memory and compute, but may provide stronger results.
Quantization
Reduced-precision model versions lower memory requirements and can improve local speed.
Context length
Long conversations and substantial notebook context can increase memory requirements.
Remember to leave memory for Python and the dataset—not only the language model.
Flexible by project
Local AI does not mean giving up the cloud
You do not have to choose one provider for every notebook. Match the model location to the project and task.
Sensitive project
Local LLM
Everyday analysis
Local or cloud model
Difficult reasoning task
Powerful cloud model
Offline environment
Local LLM
MLJAR Studio supports different provider configurations. Current documentation includes MLJAR AI, OpenAI, Ollama Local, and Ollama Cloud, giving you control over which setup fits each project.
One desktop workspace
A desktop environment for local AI + Python data analysis
MLJAR Studio puts the AI conversation and local Python execution environment in one application.
Open local files
Ask questions in English
Generate Python
Execute it locally
Inspect and edit code
Create charts and statistics
Train ML models
Save the notebook workflow
MLJAR Studio runs on your computer and is designed around local Python execution and reproducible notebooks. With Ollama configured, the language model can run locally too.
Your files
CSV · Excel · SQL
MLJAR Studio + Local LLM
Conversation and reasoning
AI Data Analyst
Generates visible Python
Python results
Tables · charts · models
Notebook
Saved and reproducible
From analysis to application
Go beyond the notebook
After cleaning or analyzing the data, turn the useful workflow into something other people can use.
Dashboard
Let someone explore metrics and charts interactively.
Chat app
Give users a conversational interface for a specific dataset or workflow.
Internal tool
Create an interface for files, cleaning, calculations, or reports.
Web app
Expose the useful part of an analysis without asking users to run a notebook.
Publish notebook workflows with Mercury
MLJAR Studio integrates with Mercury for turning notebooks into interactive dashboards, chat apps, internal tools, and web applications.
FAQ
Frequently asked questions
Can I analyze data with a local LLM?+
Yes. A local LLM can understand analytical questions and generate Python code, while Python performs calculations and transformations against the actual dataset.
Does the LLM calculate the results?+
Not necessarily. A more reliable approach for structured data analysis is to let the model generate Python and let Python perform the calculations against the real data.
Does MLJAR Studio run Python locally?+
Yes. MLJAR Studio is a desktop application and executes the analysis in the local Python environment. Generated code is visible and can be inspected and reused.
Can MLJAR Studio use a local LLM?+
Yes. MLJAR Studio currently supports and documents local LLM connections through Ollama.
Do I need to know Python?+
No. You can ask questions in plain English. Knowing Python gives you additional control, but the AI Data Analyst can generate and execute Python for you.
Which local model should I choose?+
It depends on your hardware and the complexity of the analysis. Coding quality, reasoning, memory use, speed, and the ability to complete multi-step data-analysis tasks are all important.
Can I use local AI for Excel and CSV files?+
Yes. MLJAR Studio’s AI Data Analyst can work with local structured files, including CSV and Excel, as well as SQL database connections.
Can I turn the analysis into an app?+
Yes. Notebook-based analysis in MLJAR Studio can be turned into an interactive web application using Mercury.
Start analyzing locally
Use AI to understand the question. Use Python to calculate the answer.
MLJAR Studio gives you a desktop environment for AI-assisted Python data analysis with visible code, reproducible notebooks, and support for local LLMs through Ollama.
Keep the code, data, and workflow under your control.