weights-drop

Every new open-weight model release on HuggingFace — without scrolling through 5,000 LoRAs.

A live, public dashboard that surfaces fresh base-model releases (not finetunes, not LoRAs, not requantizations), scored by week-1 download velocity. Refreshed every 2 hours from the HuggingFace public API. No auth, no accounts.

What it tracks

HuggingFace publishes 2,000–4,000 new model uploads per day, but the genuinely new base-model releases — the "DeepSeek V4 Pro" / "Qwen 3.6 Coder" / "Mistral Codestral 2" tier — number maybe 5–15 per week. They get buried under the LoRA flood. weights-drop is the page that surfaces them.

For each tracked release you get:
- author, name, parameter count, license, pipeline tag
- day-1 and week-1 download reach
- a 0–100 Velocity Score with a colored badge (🔥 frontier, ⚡ strong, 📦 solid)
- a 14-snapshot sparkline of downloads over time
- one-click link to the model on HuggingFace

How it filters

Full transparency: every constant used by the base-model filter is visible in the in-app "about the filter" page (or hit GET /weights-drop/api/about/filter directly).

A model passes the filter iff ALL of:
1. No base_model: tag (which would mark it as derived).
2. No derivative tag (lora, peft, adapter, qlora, dora, merge, mergekit).
3. No quantization tag (gguf, mlx, awq, gptq, …) — unless the author is a known first-party publisher (e.g. meta-llama, deepseek-ai, Qwen, mistralai).
4. pipeline_tag is in the accepted list (text-generation, image-text-to-text, automatic-speech-recognition, text-to-image, …).
5. Author is not on the small bot-deny list (bulk requantizers / mergers).
6. Reached activation: likes ≥ 5 OR downloads ≥ 200 within 24h, OR is first-party.

Data sources (real, no mocks)

| Source | URL | Refresh |
| --- | --- | --- |
| HF model list (newest) | https://huggingface.co/api/models?sort=createdAt&direction=-1&limit=100&full=true (× 4 pages) | every 2 hours |
| HF trending | https://huggingface.co/api/trending?type=model&limit=50 | every 2 hours |
| HF single model (drawer view) | https://huggingface.co/api/models/{author}/{name} | per-request, on demand |

If HuggingFace is unreachable during a refresh window, the failure is recorded in the fetch_errors table and the dashboard's "last refresh" footer reflects the last successful snapshot. No fallback data is synthesized.

Run locally

npm install
cp .env.example .env       # then edit PORT / DB_PATH if needed
npm start

Open <http://localhost:4789/weights-drop/>.

The first refresh runs ~1s after boot if the DB is empty or the last snapshot is older than 2h. You'll see real rows appear within a few seconds.

Endpoints

All endpoints are public, mounted under /weights-drop.

• GET / — SPA
• GET /api/drops?range=24h|7d|30d|all&pipeline=&org=&minScore=&sort=score|created|downloads&limit=&offset= — board JSON
• GET /api/drop/:author/:name — single-model detail including snapshots
• GET /api/stats — top-of-page stat cards
• GET /api/feed.json — JSON Feed 1.1 of the last 50 drops
• GET /api/feed.rss — RSS 2.0 variant
• GET /api/about/filter — full filter & scoring configuration
• POST /refresh — manual refresh (rate-limited 1 per 30s per IP)
• GET /health — liveness probe

Stack

• Node.js 20+, Express 4
• better-sqlite3 (WAL)
• node-cron (0 /2 fetch, 30 /6 re-score, 15 4 * housekeeping)
• helmet + compression + express-rate-limit
• Vanilla JS SPA (no framework), dark theme, English labels

License

MIT.