Replicate — AI Glossary

Reference details

Related terms

• Modal
• RunPod
• Serverless Inference

How it works

1. Each model is packaged as a container with a declared input and output schema, so running it needs no knowledge of its internals.
2. Versions are immutable: a version identifier always refers to the same image, which is what makes a prediction reproducible.
3. Predictions are asynchronous by default — a job is created and polled or webhooked — because many hosted models take seconds to minutes.
4. Billing is per unit of compute consumed by the prediction, so an idle model costs nothing and a slow one costs more.
5. Anyone can publish a model, which is why the catalogue is broad and its quality uneven compared with a curated inference service.

Example: Immutable versions are the reproducibility mechanism

A pipeline pins a model by version identifier. The author publishes an improved version; the pipeline keeps producing identical outputs because the identifier still resolves to the original image. Pointing at the model by name instead would have picked up the change silently, and output drift would have had no corresponding commit. This is the same discipline as pinning a dependency, with the difference that the behaviour change here is statistical rather than a build error.

Commonly confused with

Differs from Together and Fireworks in unit of deployment: models are versioned containers you can publish and run, not a fixed catalogue behind one API. That makes it the option when the model you need is not on a major provider's list, and the reason cold starts are a real consideration — a container that has not run recently has to be brought up before it answers.

When to use it

Reach for it when:

• Running a wide range of community or multimodal models without packaging them yourself.
• Reproducibility requires pinning an immutable model version.
• Workloads are intermittent, so per-prediction billing beats reserved capacity.

Reach for something else when:

• Low, predictable latency is required — cold starts and queueing are part of the model.
• You need a curated, SLA-backed catalogue of frontier text models.
• Sustained high volume, where dedicated capacity is cheaper.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Replicate — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path