Get the latest setups
THE NUMBERS, EXPLAINED

Age. Adoption. Preference.

Freshness tells you how recently a model arrived. Popularity shows activity around its weights. Arena shows how people rated it against other models. Use them alongside hardware fit and the setup guide.

Freshness

A new release starts at 100. The score declines with age, reaching 50 after six months and roughly 25 after a year.

round(100 × 2^(−age in days / 180))

We use the catalog’s model release date. Where that is unavailable, we show the exact weight repository’s creation date and label it repository age. A recent quantization upload can therefore have a young repository without being a new base model.

README edits, source checks, and guide repairs never make a model look new. Missing or future dates stay unscored.

Popularity

Hugging Face downloads over the past 30 days account for 80% of the score. Lifetime repository likes account for 20%.

Both counts use a logarithmic scale: a move from 1,000 to 10,000 downloads matters as much as a move from 100,000 to 1 million. Downloads cap at 10 million; likes cap at 100,000.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

If likes are unavailable, downloads carry the entire score. Missing downloads stay unscored. A reported zero is a real value; missing data is not zero.

What counts as a download?

Hugging Face counts requests to selected model files under its library-specific counting rules. Downloads are not unique people or installations, and different model formats may have different counting behavior. Comparisons within a similar model type are more useful than treating the score as a universal leaderboard. Hugging Face download-counting documentation ↗

Arena · Open Source rank

A position within Arena’s Open Source filter, not a score out of 100. Lower is better. Each rank names its category: Text, Coding, Agents, Vision, Documents, or a specific image/video task. We take the complete Overall category in Arena’s official export, apply its Open Source license filter, preserve the exported rank order, and number the remaining entries. Text, Vision, and Documents use Arena’s default style-controlled exports.

Arena’s filter includes some custom and noncommercial licenses. It is not a guarantee of OSI compliance or permission for every use. Rankings reflect human preferences, not download counts, local speed, or hardware suitability. We do not combine category ranks or mix them into popularity. The model’s main category determines which badge appears first; the expanded view lists its other matched categories.

The badge identifies the exact Arena model entry and export date. The live website and export can update at different times. Equal rounded scores and confidence ranges are not treated as tied ranks. Data: Arena’s official leaderboard dataset ↗, licensed under CC BY 4.0 ↗. YouRunAI calculates filtered positions from that export. Arena’s ranking method ↗

Exact models, not similar names

Arena matching checks publisher and checkpoint identity. We do not drop Flash, Pro, Max, thinking, or revision dates to guess a match. Modified, distilled, and uncensored variants do not inherit a parent’s ranking. A confirmed packaging option may show Original model ranking; this is not a benchmark of that quantization. New catalog models automatically receive these rules. If Arena has not ranked an exact model, no Arena badge appears. We never invent a rank or substitute zero.

Updated daily, without generating content

The daily Vercel refresh runs at 13:20 UTC, retrieving Hugging Face metadata and ten Arena boards. It uses no AI tokens and does not publish content. Pages read the saved snapshots, so browsing never triggers another leaderboard download. Freshness is recalculated when the page loads. Popularity and Arena show when their sources were checked; observations older than 72 hours are marked “Refresh due.” If a source fails, we retain its last complete observation. An incomplete leaderboard never replaces a complete one. Arena may not publish a new board every day: an older export date is different from an overdue source check.

The right model, including in a workflow

Workflow scores describe the linked model, not the age of the guide. Package options use explicit model links or exact source repositories. A separately hosted uncensored or quantized variant may have different activity from its parent. We do not guess a match from a shared family name.

Filters, with your choices preserved

Filter by model age or a minimum popularity score in Models, Workflows, or Start Here. A workflow matches when at least one of its models meets both filters. Models missing the required data are excluded only when that filter is active. The usual unfiltered view still includes them.

In Models, choose an Arena category to filter by Open Source rank: ranked only, Top 5, Top 10, Top 25, Top 50, or models not ranked in that category. Arena sorting puts #1 first and models without a rank last. Rankings from different categories are never combined. Within a model family, the same version must satisfy your rank, hardware, and other filters; its card shows the selected category’s rank. The most recently saved Arena observation is used, with overdue checks still labeled “Refresh due.”

Start Here’s featured recommendations remain subject to editorial approval. Sorting the browse list by scores does not replace a featured model or override hardware compatibility.

A score is context, not a verdict. An older model can still be your best choice. A popular model can still be too large for your machine. Follow the documented package, hardware guidance, and workflow.

Explore models