Far right and small = published days ago and already liked, before the download counter can move. That corner is invisible to any table ranked by downloads. Third-party quantizations are excluded here and kept in the catalog below; the likes axis is logarithmic.

Source: Hugging Face Hub — API documentation · Data as of 2026-08-30 17:07 UTC · source code

LLM Radar is an open dashboard that consolidates the large-language-model landscape into one continuously-updated view, split into two catalogs:

• Hosted — commercial models you call through an API. Sourced from the public OpenRouter /api/v1/models endpoint: pricing per million tokens (input/output), context-window size, modalities, and release date.
• Open-weights — models you can download and self-host. Sourced from the public Hugging Face Hub model listing: 30-day downloads, community likes, license, gated status, and type.

How the open-weights universe is chosen. Two decisions shape what appears, and both exist to stop the catalog from lagging the market:

• Not only text. Open models are no longer just text-generation; a new multimodal release lands under image-text-to-text and would be invisible to a text-only query. Each tracked type is queried separately and labelled.
• Not only downloads. The 30-day download counter takes weeks to move, so a ranking built on it can only show what was already popular — a model published two days ago sits at zero however important it is. Each type is therefore also pulled by the Hub’s trending signal, which surfaces a release the day after it lands. Both lists are merged.

The data refreshes daily via a GitHub Actions cron that re-fetches both public APIs, re-renders the dashboard, and redeploys — no server, no backend, no API key.

• Context window (K) — the maximum number of tokens (prompt + response) a model can process in one call, in thousands. A 200K context ≈ ~150,000 words.
• $/1M in · $/1M out — price in US dollars per one million input / output tokens, as published by the provider on OpenRouter.
• Permissive license — an open-weights license (Apache-2.0, MIT, and similar) that allows commercial use and redistribution without acceptance gates. The opposite is a restricted or research-only license.
• Gated model — an open-weights model whose download requires accepting terms or requesting access. Relevant for data-residency and commercial-use decisions.
• Cheapest capable — lowest input price among models with ≥100K context, i.e. the cheapest model still large enough for real document / agent workloads.
• Type — what a model can be fed. On the hosted side it is derived from the input modalities OpenRouter publishes (text only · image · document + image · audio / video); on the open-weights side it is the Hub pipeline tag (text · multimodal · vision / OCR). It is read from the API, never assigned by hand.
• Just landed — models published in the last 45 days, ranked by likes. Ranking recent arrivals by downloads would return an empty table: the counter has not had time to move.
• Third-party quantization — a repackaging of someone else’s weights (GGUF, AWQ, MLX, 4-bit and similar). Real downloads, but not a new model, and numerous enough to crowd out the originals. Kept in the searchable catalog, excluded from the charts and the ranked tables.
• Open weights available (hosted chart) — the hosted model has a published Hugging Face repository, so the same weights can be self-hosted. Shown as a filled marker; closed models are hollow.

Pipeline: OpenRouter API + Hugging Face API → Python (pandas, great-tables) → Quarto dashboard → GitHub Pages, refreshed daily by GitHub Actions. Fully open-source and reproducible — the entire pipeline is in the source repository.

LLM Radar is an independent project and is not affiliated with OpenRouter or Hugging Face. Figures reflect what those public APIs report at fetch time; always confirm pricing and licenses against the provider before relying on them.

To cite: LLM Radar — the full LLM landscape (hosted pricing & context plus open-weights adoption & licenses), by Ronald Mego. https://ronaldmego.github.io/llm-radar