Model zoo

A model zoo is the platform-hosted collection of pre-trained models you can browse, download, fine-tune, or serve — Hugging Face Hub (the dominant 2026 model zoo), Replicate, Civitai, ModelScope are leaders.

Model zoos solve the discoverability problem: with thousands of open-weight LLMs + image models + speech models, where do you find the right one? Hugging Face Hub dominates with 1M+ models, datasets, leaderboards, and one-line Python download. Replicate hosts inference-ready endpoints. Civitai specializes in Stable Diffusion checkpoints + LoRAs. ModelScope is Alibaba's Chinese-focused alternative. Engineering touchpoints: model cards (HF standardized format documenting use, limits, biases), versioning + revisions, gated access (some models require approval — Llama, some Mistral tiers), Spaces (HF's hosted demo environment). Production patterns: pin a specific revision (not main), audit the model card, mirror weights to your own CDN for production reliability.

When to use model zoo

• Discovering open-weight models.
• Hosting your own fine-tunes for community.

Common mistakes

• Loading `main` revision in prod — breaks if author force-pushes.
• Not auditing the model card — biases / limits get missed.

FAQ

Related terms

• Open-weight model — An open-weight model has publicly released weights downloadable + runnable by anyone — Llama, Mistral, Qwen, DeepSeek, Flux Schnell / Dev, Stable Diffusion are 2026 open-weight families. Differs from open source (which would include training code + data).
• Model card — A model card is structured documentation accompanying a released model — what it does, what it was trained on, its evaluation results, intended uses, and known limitations.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/model-zoo.md.