---
title: Open LLM Distribution Leaderboard
emoji: 📊
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
short_description: Daily open LLM leaderboard.
---

# Open LLM Distribution Leaderboard

This Space serves a daily HTML leaderboard for popular open LLM candidates on Hugging Face. A scheduled refresh job stores mutable data in a Hugging Face bucket instead of using the Space repo as the backing store.

## Data Flow

- Read public Hugging Face model metadata for the configured model creators.
- Add live Inference Providers router metadata when available.
- Discover popular third-party runtime artifacts from global GGUF, quantized, FP8, and MLX searches.
- Filter for LLM-ish text, coding, multimodal, and special-architecture model candidates.
- Group runtime artifacts by normalized model identity, explicit packaging metadata, and conservative lineage slug evidence.
- Write `latest/` and timestamped `snapshots/` into the mounted bucket.
- Serve `latest/index.html` from the Space, with a default one-year age filter in the UI.

## Bucket

The production bucket is expected at:

`hf://buckets/osolmaz/leaderboard`

The Space mounts it at `/data`.

If the mount is not ready when the web process starts, the app fetches `latest/` from `OPEN_MODEL_BUCKET_URI`, which defaults to the same bucket, into the local fallback data directory.

Source links:

- Space: <https://huggingface.co/spaces/osolmaz/leaderboard>
- Bucket: <https://huggingface.co/buckets/osolmaz/leaderboard>

## Daily Refresh

Production refresh should run from a Hugging Face Scheduled Job that mounts:

- the Space repo at `/workspace`
- the bucket at `/data`

and runs:

```bash
python /workspace/refresh_once.py
```

That external job path requires `job.write` permission and Jobs billing/credits for the account or organization that owns the job.

The Space web process should not run automatic refreshes in production:

```text
OPEN_MODEL_REFRESH_ON_STARTUP=0
OPEN_MODEL_ENABLE_SCHEDULER=0
```

Discovery is enabled by default during refreshes. Useful local knobs:

```text
OPEN_MODEL_ENABLE_DISCOVERY=0
OPEN_MODEL_DISCOVERY_MAX_RESULTS_PER_QUERY=500
OPEN_MODEL_RELATED_DISCOVERY_SEED_LIMIT=12
```

## Local Smoke Test

```bash
python leaderboard_core.py --output-dir /tmp/leaderboard-data --limit-per-creator 1
python -m http.server 7861 --directory /tmp/leaderboard-data/latest
```