Hosted inference

Hosted inference is the right use case when the model creator does not run a paid first-party API. Meta is the canonical example today: Llama 4 Scout and Llama 4 Maverick are open-weights releases with no Meta first-party API; inference is delivered by third-party hosting platforms (Groq, Together AI) at rates each platform sets independently.

The catalogue separates two distinct claims here:

• Hosted availability is a stable identity fact — the platform exposes the model under a specific hosted model ID (e.g. Groq's meta-llama/llama-4-scout-17b-16e-instruct). Availability rarely changes and is recorded in lib/hosted-availability.ts.
• Hosted pricing is a volatile reference value — the platform's published per-token rate on a specific date. Hosted pricing carries a freshness chip and a high-volatility tag because hosting platforms re-price frequently. See /research/api-pricing-methodology for the full distinction.

Two read-traps to avoid:

• A hosting platform is not the model creator. Groq does not become "the maker of Llama 4 Scout" by exposing it. The catalogue refuses to attribute creator status to a hosting platform; the integrity guards block any such drift.
• Hosted pricing is not creator pricing. Groq's Llama 4 Scout rate reflects Groq's pricing decision — not Meta's. WebmasterID Models does not rank hosting platforms by price, and you should not infer that Together's rate is "cheaper" or "more expensive" in a meaningful sense without a workload-specific projection.

The shortlist below collects models with verified hosted availability records. Each row links to the hosting platform's provider page, where the availability sidebar + hosted pricing references render side-by-side with the verification + freshness state. Re-verify against the platform's own pricing page before any procurement decision.