Key Terms

Vocabulary used throughout this guide and inside the Foundry dashboard.

Term	Definition
Model HUB	The catalog of every model available to your organization for testing and deployment, whether imported from Hugging Face or provided by Multiverse.
Model version	A specific, immutable snapshot of a model. Each version has its own status (e.g. ready) and version ID, and is deployed independently.
Deployment	A running, hosted instance of a model version, backed by allocated CPU, GPU, memory, and replicas. This is what "hosting" a model means in Foundry.
Playground	A sandbox chat interface for testing a model's responses, speed, and cost before wiring it into a real workflow or deployment.
vLLM	The inference engine/container image that serves deployed models. Each deployment specifies which vLLM version to run.
Replica	One running copy of a deployment. Increasing replicas adds capacity for concurrent requests; it doesn't change model quality.
CompactifAI	Multiverse Computing's model compression product. It shrinks a model's size while trying to preserve as much quality as possible.
Compression job	A workflow that compresses one model to a target compression rate for a specific use case, then runs healing steps to recover quality.
Compression rate	The target percentage size/parameter reduction for a compression job. Higher rates save more compute but carry more quality risk.
Healing steps	Post-compression recovery stages — Recovery, Consolidation, Adaptation, and Localization — that restore model quality lost during compression.
Compression speed	How many GPUs are allocated to a compression job. Faster settings finish sooner but use more compute at once.
Use case	The intended usage pattern for a model (e.g. Agentic / Tool calling), used to tune how a compression job is run.