Best overall
gpt-oss:120b
9.87/10 · std 0.45
23 real Python workflows
Best Local LLM for Data Analysis 2026 Benchmarks
Most lists rank models on trivia, math puzzles, or chat quality. We tested the job that matters here: exploring real datasets, writing working pandas code, recovering from errors, and explaining the result.
Six models, the same multi-step workflows, public notebooks, and one shared scoring rubric.
MLJAR benchmark leader
Average across 23 published scenarios
gpt-oss:120b
9.87
out of 10
0.45
score deviation
Open
runs via Ollama
The result
Quick answer
Best balanced
glm-5.1
9.48/10
Strong 31b option
gemma4:31b
9.30/10
Weakest on complexity
qwen3.5:397b
6.33/10 on complex tasks
Also worth testing: GLM-5.2, DeepSeek V4 Flash and Pro, Kimi K3, and Qwen 3.6 arrived after our April benchmark and are not included in the MLJAR ranking yet.
A different skill profile
What “best” means for data analysis, not chat
A model that writes a clever poem or passes a trivia benchmark is not necessarily reliable inside a long analytical workflow.
- 1
Understand the analytical intent behind the request
- 2
Generate Python that runs against real data
- 3
Use pandas, NumPy, and plotting libraries correctly
- 4
Interpret output from the previous step
- 5
Recover from errors instead of repeating them
- 6
Keep a multi-step analysis coherent
- 7
Explain results so a non-programmer can act on them
That is why we use a domain-specific local LLM benchmark rather than borrowing a general chat leaderboard.
First-party results
Our benchmark: real Python tasks, not trivia
Six models completed 23 published scenarios spanning exploratory analysis, time series, machine learning, finance, statistics, and NLP.
Each scenario is a sequence of prompts. Every run is scored for task completion, execution correctness, output quality, reasoning, and reliability. The complete notebook for every run is public.
| Model | Average score | Open weights | Runs via Ollama |
|---|---|---|---|
| gpt-oss:120b | 9.87 / 10 | Yes | Yes |
| gpt-5.4 | 9.65 / 10 | No — cloud only | No — cloud only |
| glm-5.1 | 9.48 / 10 | Yes | Yes |
| gemma4:31b | 9.30 / 10 | Yes | Yes |
| qwen3-coder-next | 9.04 / 10 | Yes | Yes |
| qwen3.5:397b | 8.43 / 10 | Yes | Yes |
An open-weight model finished first
gpt-oss:120b outscored gpt-5.4, the cloud-only model in this set, and produced the smallest score deviation across all 23 workflows.
Model size did not ensure reliability
qwen3.5:397b averaged 6.33/10 on complex scenarios. On the Value-at-Risk workflow, every other model scored 9 or 10 while its run scored 2/10.
Candidates for the next run
Newer models worth testing
The models below appeared after our April benchmark. Independent coding results help form a shortlist, but they are not MLJAR data-analysis scores.
SWE-Bench Verified
Resolved GitHub issues · higher is better
DeepSeek V4 Pro
80.6%
MiniMax M3
80.5%
Kimi K2.6
80.2%
DeepSeek V4 Flash
79.0%
Kimi K3
76.8%
Kimi K2.5
76.8%
DeepSeek V4 Pro
80.6% SWE-Bench Verified and 93.5% on LiveCodeBench. A Mixture-of-Experts model built for efficient reasoning over very long context.
MiniMax M3
80.5% SWE-Bench Verified. A newer entrant aimed squarely at coding and agentic workflows.
Kimi K2.6
80.2% SWE-Bench Verified and strong computer-use performance. It is the newest step in Kimi’s agentic model line.
DeepSeek V4 Flash
79% SWE-Bench Verified. The lighter sibling of V4 Pro is positioned for faster inference while retaining strong coding performance.
Kimi K3
76.8% SWE-Bench Verified in the comparison data, with a broader strength profile that emphasizes general reasoning and terminal tasks.
Kimi K2.5
76.8% SWE-Bench Verified, one step behind the newer K2.6 and K3 releases.
GLM-5.2
No public SWE-Bench Verified score in the cited tracker yet, but it is positioned specifically for long-horizon agentic tasks.
Qwen 3.6
Worth watching, but public SWE-Bench claims vary too widely by model variant and evaluation harness to cite one responsible number yet.
Mistral Medium 3.5, Nemotron 3, and Granite 4.1 are also coding- or agentic-focused candidates, but we did not find sufficiently comparable public SWE-Bench or LiveCodeBench figures to place them in this chart. Qwen 3.6 is omitted for the same reason: public claims vary materially by variant and harness.
Supporting evidence
What independent coding benchmarks show
Python data analysis depends on code generation, but coding leaderboards still measure something different from an iterative workflow with a dataset.
Treat these scores as context, not a substitute for domain-specific testing. Evaluation harnesses, languages, tools, and task shape all change the result.
Aider Polyglot
41.8%
gpt-oss:120b
225 difficult exercises across C++, Go, Java, JavaScript, Python, and Rust.
LiveCodeBench leader
93.5%
DeepSeek V4 Pro
DeepSeek V4 Flash follows at 91.6%; Kimi K2 Thinking scores 83.1%.
LiveCodeBench
69%
gpt-oss:120b and 20b
Solid code generation, but behind the newer DeepSeek V4 family on this test.
MLJAR workflows
9.87/10
gpt-oss:120b
The same model leads when tested on iterative Python data analysis.
Read together with our results, the newer DeepSeek and Kimi releases are strong candidates for the next MLJAR data-analysis benchmark run.
A model you can actually run
Pick for your hardware, not only the leaderboard
Most model families come in several sizes. Smaller variants need less memory, but may trade away some of the reliability shown above.
31b tested
Gemma 4
e2b → 31b
Small variants fit lighter machines; the 31b model produced the 9.30/10 result in our benchmark.
397b tested previously
Qwen 3.5
0.8b → 122b
The family covers a wide range of hardware, but smaller variants may trade away workflow reliability.
120b tested
gpt-oss
20b or 120b
The 120b model led our benchmark; the smaller 20b variant is the more accessible local option.
Three steps
Try any model in MLJAR Studio
Once Ollama serves the model, MLJAR Studio can use it for the same visible, notebook-based Python workflow.
- 1Install Ollama and pull the model.
- 2Open AI provider settings and select Ollama.
- 3Enter the model name, test the connection, and save.
terminal
ollama pull gpt-oss:120b
ollama run gpt-oss:120bPull the model once, test it locally, then select the same model name in MLJAR Studio.
FAQ
Frequently asked questions
What is the best local LLM for data analysis right now?+
In our benchmark of real Python data-analysis workflows, gpt-oss:120b scored highest at 9.87/10 and was also the most consistent model tested.
Can an open-weight model really beat closed models on data analysis?+
In our testing, yes. gpt-oss:120b, which runs locally, outscored gpt-5.4, the cloud-only model in the same test set.
Do I need a GPU to run these models?+
It depends on model size. Smaller few-billion-parameter variants can run on CPU-only laptops, while 120b-class models benefit significantly from GPU or unified memory.
Is a coding benchmark score the same as a data-analysis score?+
No. Coding benchmarks test code generation or issue resolution. The MLJAR benchmark tests iterative, multi-step analysis against real datasets, which is more directly relevant here.
How often is this comparison updated?+
Capable open-weight models appear frequently. We update this page as models enter our benchmark; the newer-model shortlist tracks candidates that MLJAR has not tested yet.
Analyze your data with the model that actually performs
Do not take a chat leaderboard’s word for it. Inspect every prompt, generated line of Python, output, and chart in the full MLJAR benchmark.