Decrypting snapshot
Loading data/models.json …
▪ View 01 — trending
Trending by modality
Each band is the TOP 5 by Hugging Face trendingScore at snapshot time, ranked within its own pipeline_tag. Ties break on 30-day downloads.
▪ View 02 — index
Model index
The full snapshot. Search matches model id, author and tags; results are paginated at 50 rows.
| # | Model | Author | Pipeline | DL 30d | DL all-time | Likes | Trend | License | Updated |
|---|
▪ View 03 — providers
Provider leaderboard
Records grouped by author, ranked by total 30-day downloads. All-time totals are shown alongside, since the two orderings differ.
| # | Provider | Models | DL 30d | DL all-time | Likes | Top pipeline |
|---|
▪ View 04 — hardware
GPU compatibility
Weight-memory estimates for the text-generation models whose parameter count is published. Estimates cover model weights plus a flat runtime overhead — they are not a benchmark.
Estimation formula
- FP16
- 2.00 B/param
- Q8
- 1.07 B/param
- Q6
- 0.82 B/param
- Q5
- 0.72 B/param
- Q4
- 0.60 B/param
FP16 = 2.00, Q8 ≈ 1.07 and Q4 ≈ 0.60 bytes per parameter are the reference values; Q6 ≈ 0.82 and Q5 ≈ 0.72 are interpolated between them for the mid-range quantisations. The flat 1.5 GB covers CUDA context, activations and a short KV cache — long-context work needs more, and MoE checkpoints load all experts into VRAM even though only a few are active per token.