Methodology

Every number on this site is reproducible. This page states exactly where each one comes from.

Download data

We query the public Hugging Face Hub API nightly and store a snapshot of every catalog model's trailing-30-day downloads, all-time downloads and likes at 04:00 UTC. Hugging Face shows only the current trailing-30-day figure; the day-by-day history — trend charts, 7-day growth rankings, crossover dates — exists because we record it ourselves, daily, since 28 Jul 2026. The catalog covers roughly the top 1,000 models, admitted per task category by fixed quotas so a text-generation giant never crowds out the top speech or embedding models. Known CI/test repositories with inflated download counters are excluded.

VRAM estimates

est_vram_gb = file_size_gb × 1.1
            + kv_cache(context, params)   # 0.15 GB per B params at 8K ctx, linear in ctx
            + 0.5 GB runtime overhead

For GGUF files the quantization is already reflected in the file size, so the file size is the input. For unquantized checkpoints we use parameter count × 2 bytes (fp16). The KV-cache term is a deliberate simplification — a linear approximation calibrated at 8K context — because the honest alternative (per-architecture head math) changes the answer by less than the runtime variance between llama.cpp and vLLM. Verdicts against a GPU: comfortable below 82% of usable memory, tight up to 100%, won't fit above. Apple-silicon presets count 75% of unified memory as usable. These are estimates, not benchmarks; measured tokens/sec is a future, crowdsourced feature.

Licences

The "commercial use" badge is a three-way classification of the repository's declared licence: permissive licences (Apache-2.0, MIT, BSD, CC-BY…) map to allowed; non-commercial licences (CC-BY-NC and variants) to not allowed; community licences with usage conditions (Llama, Gemma, OpenRAIL, FLUX…) to conditional. It is a signpost, not legal advice — always read the licence text linked on the model page.

Verdicts and prompt outputs

Comparison verdicts are generated by a language model from structured data only — the two models' specs, benchmark results, licence classes, VRAM estimates and download trends. The generator is not allowed to use its own knowledge about the models, which prevents confident hallucination. Prompt-gallery outputs are real completions produced once through Hugging Face inference providers, cached permanently, and never edited or cherry-picked. Both carry an "auto-generated" label wherever they appear.

Questions or corrections: the catalog is data-driven end to end, so if a number looks wrong the fastest fix is an issue with the model's /api/v1/models/[slug] payload attached (example: qwen3-8b).