# Distributed Inference
**Entity class:** Distributed-computing concept
**As of:** November 17, 2026
## Definition
Distributed inference runs a trained model’s forward computation across more than one device or places complete model replicas on several hosts. [[wiki/Petals|Petals]] distributed layer blocks over the public internet; [[wiki/exo|exo]], [[wiki/prima.cpp|prima.cpp]] and [[wiki/TPI-LLM|TPI-LLM]] address local or constrained-device deployments. A complete replica at each capable host favors independent continuity, while a model split by layers makes each token depend on inter-host communication. Latency, bandwidth, memory and failure patterns determine the better arrangement for each workload.
## Relationships
- **ecosystem router:** [[wiki/Distributed Compute Ecosystem|Distributed Compute Ecosystem]].
- **worked example:** [[wiki/Petals|Petals]].
- **source analysis:** [[articles/Embedded Super-Spreader Events and Muse|Embedded Super-Spreader Events and Muse]].
## Sources
- [Distributed Inference: primary project record or technical account](https://arxiv.org/abs/2209.01188).