Open LLM Leaderboard Scraper

Open LLM Leaderboard Scraper

Scrape the Open LLM Leaderboard for model rankings, benchmark scores, and parameter counts, up to a million models per run. Every model comes with its average score, ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K, and parameter count. No login or API key. Export to CSV, JSON, Excel, or XML.

The Open LLM Leaderboard on Hugging Face tracks thousands of large language models, but its web interface is hard to query and compare at scale. This Actor reads the public leaderboard directly, filters by model name, precision, or parameter count, and returns each match in one fixed schema. No API key, no rate limits, no manual copying.

Who uses it	What they scrape Open LLM Leaderboard for
AI researchers	Which open models currently lead on specific benchmarks like MMLU or GSM8K
ML engineers	Shortlist candidate models by parameter count and average score for fine-tuning
Product managers	Track how a competitor's model moves up or down the leaderboard over time
Data analysts	Build a local dataset of model performance for custom ranking or visualization
Tech journalists	Pull the latest top-10 open models with scores for a news article

What it does

This Actor collects Open LLM Leaderboard entries by model name search, precision, or parameter range, and returns each model as a flat row with its benchmark scores and metadata.

• 🔍 Model name search: case-insensitive substring match on the model name, so 'llama' finds all Llama variants.
• 📊 Benchmark filters: filter by precision (float16, bfloat16, float32, 8bit, 4bit, GPTQ) and parameter count range.
• 📈 Sortable results: sort by average, ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K, or parameter count.
• 📦 Bulk export: fetch up to 1,000,000 models on paid plans, 10 on free preview, in CSV, JSON, Excel, or XML.

Results export to CSV, JSON, Excel, or XML, or straight from the API.

What you can do with Open LLM Leaderboard data

📈 Track model rankings over time.

An AI researcher runs the Actor weekly with no filters, sorts by average, and logs the top 50 models to spot new entrants and ranking shifts.

🔍 Shortlist models for fine-tuning.

An ML engineer sets maxParams to 13 and precision to bfloat16, then sorts by MMLU to find small, capable models that fit on a single GPU.

📊 Compare open models on a specific benchmark.

A product manager searches for 'mistral' and sorts by GSM8K to see which Mistral variant scores highest on math reasoning before choosing an API provider.

📚 Build a local leaderboard dataset.

A data analyst pulls all models with minParams 7 and exports to CSV, then joins the data with license information for an internal dashboard.

📰 Report on the state of open LLMs.

A tech journalist runs the Actor with maxItems 10 and sortBy average to get the current top models and their scores for a news article.

Why choose this scraper

	What you get
No API key	Reads the public leaderboard directly, no Hugging Face token or registration needed.
One flat row per model	Every model returns the same fields: name, average, ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K, and parameters.
Filter before you download	Search by model name, precision, and parameter count to get only the models you care about.
Sort by any benchmark	Order results by average or any individual score to see leaders at a glance.
Scale to a million models	Paid plans can pull the entire leaderboard in one run; free preview returns 10 models.

How it compares

No other Store actor targets Open LLM Leaderboard the same way, so the honest comparison is with the alternatives teams actually weigh.

	Open LLM Leaderboard Scraper	Build it in-house	By hand
Setup	Run it now, zero config	Days of engineering	None, but hours per pull
When Open LLM Leaderboard changes	Maintained for you	You fix it	You re-learn the page
Proxies, retries, anti-bot	Built in	Your problem	Browser only
Output	Fixed JSON schema, CSV/Excel export	Whatever you build	Copy-paste
Cost	Pay per result	Engineering time	Analyst hours

Configure the run

Drive the Actor with a model name search, precision, and parameter range, and sort the results by any benchmark score. Filters run client-side after fetching the leaderboard, so only matching models reach your dataset. The Input tab lists every parameter.

A first run with the defaults:

{
  "maxItems": 10
}

A larger pull:

{
  "maxItems": 200
}

Pricing

Pay-per-result: $0.001 per result collected. You pay only for the results written to your dataset.

Results collected	Approximate cost
100 results	$0.10
1,000 results	$1.00
10,000 results	$10.00

New Apify accounts start with $5 in free credit.

Free users

Free-plan runs return up to 10 results as a preview. Upgrade your Apify plan to collect up to 1,000,000 results per run.

Run it

1. Create a free Apify account with $5 in credit.
2. Open the Open LLM Leaderboard Scraper.
3. Set your inputs and any filters, then click Start.
4. Export the results as CSV, Excel, JSON, or XML from the Dataset tab.

Run it programmatically through the Apify API (run-sync-get-dataset-items) or the ApifyClient for JavaScript and Python.

Use with AI agents (MCP)

Give an AI agent live access to Open LLM Leaderboard through the Model Context Protocol. Add the Actor to Claude, Cursor, or any MCP client:

$claude mcp add --transport http apify "https://mcp.apify.com?tools=parseforge/openllm-leaderboard-scraper"

Then prompt it in plain language to run the scraper and read back the results.

Troubleshooting

Why am I getting no results?

Your filters may be too restrictive. Remove the search term, widen the parameter range, or set precision to blank, then run again. The leaderboard may also have changed since your last run.

Why did I only get 10 models?

Free preview is limited to 10 models. Upgrade to a paid Apify plan and set maxItems higher to fetch more.

Why are some models missing from the results?

The Actor reads the public leaderboard, which may not include every model ever submitted. Also check your precision filter: a model reported as float16 will not appear if you filter for bfloat16.

Why is the sort order not what I expected?

Make sure sortBy is set to the field you want. The default is average, but you can choose any benchmark or parameter count. Note that sorting is applied after filtering.

Why does the run take a long time?

Fetching a large number of models takes time. Reduce maxItems or narrow your filters to speed up the run.

FAQ

Question	Answer
Do I need a Hugging Face API key?	No. The Actor reads the public Open LLM Leaderboard web page directly, so no registration or token is required.
How many models can I scrape in one run?	Free preview is limited to 10 models. Paid plans can set maxItems up to 1,000,000, which covers the entire leaderboard.
Can I filter by model size?	Yes. Use minParams and maxParams to set a parameter count range in billions. For example, minParams 7 and maxParams 13 returns models between 7B and 13B parameters.
What does the precision filter do?	It filters models by the precision reported on the leaderboard: float16, bfloat16, float32, 8bit, 4bit, or GPTQ. This is useful if you only want models that run in a certain quantization.
Can I sort the results?	Yes. Use sortBy to order by average, ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K, or parameter count. The default is average.
What output formats are supported?	The Actor exports to CSV, JSON, Excel, and XML. You can choose the format when you set up the run in Apify.
Does the Actor return all benchmark scores?	Yes. Each model row includes the average score plus ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, and GSM8K, along with the parameter count.
Is the data live?	The Actor fetches the current leaderboard at the time of the run, so the data is as fresh as the public page.
Can I search for a specific model name?	Yes. The search input does a case-insensitive substring match, so 'llama' returns Llama 2, Llama 3, CodeLlama, and any other model with 'llama' in the name.
What if I get no results?	Check your filters. The search term may not match any model name, or the precision and parameter range may be too narrow. Try removing filters one at a time.

Related actors

Browse the full ParseForge collection for more scrapers.

🆘 Need help? Email parseforge@protonmail.com with your run ID, your input, and what you expected.

⚠️ Disclaimer. This Actor is unofficial and is not affiliated with, endorsed by, or sponsored by Hugging Face, Inc. It collects only publicly available data. You are responsible for using the collected data in compliance with the source's terms of service and applicable data-protection laws, including GDPR, CCPA, and PIPL. Do not use it to collect personal data unlawfully.