A registry first, an inference vendor second

Technology

Hugging Face is where open models live: weights, datasets, tokenizers, and the leaderboards that let you compare them. That is the durable part. The inference products layered on top are useful, and they are not the reason the Hub matters.

We run frontier models through Bedrock, Vertex AI, and the Vercel AI Gateway. The Hub is where we look when a job wants a small task specific model instead: an embedding model, a reranker, a classifier, a transcription model. Different question, different shelf.

Monogram logo

The Hub

What a model registry is actually for

  1. Small specialized models beat big general ones

    Embedding, reranking, classification, transcription: narrow jobs where a model trained for the task outperforms a frontier chat model and costs a fraction per call. The Hub is where those models are versioned, licensed, and benchmarked in one place, and that cataloguing is the part that is hard to replace.

  2. Open weights is not one license

    Repositories on the Hub carry everything from Apache 2.0 to gated access to community licenses with usage thresholds and attribution requirements attached. Open weights describes the download, not your right to ship it. Read the license before the model reaches a build, not during legal review afterward.

  3. Self hosting trades per token for per hour

    A hosted API bills per request and costs nothing while idle. A dedicated endpoint bills for the GPU whether traffic arrives or not. Below sustained steady load the managed API is cheaper, and the honest case for self hosting is data residency or tail latency, not price.

The decision the Hub actually forces is where a model runs, not which one wins a benchmark. Weights you can download are weights you can pin, audit, and deploy inside a client's own account, which counts for more in regulated work than a place on a leaderboard.

Hugging Face has not been the center of a build we have shipped. The models we run in production reach us through Bedrock, Vertex AI, and the Vercel AI Gateway, and the Hub is where we go to evaluate the alternatives before that decision gets made. Start a conversation.