> Canonical page: https://posttrainllm.com/inspiration/teale-petals

A note of thanks · Scale boundary

# teale and Petals

These decentralized inference projects make the latency cost of sharding a model across machines tangible. The comparison helps explain why a single-Mac specialist is a different system.

01 · The idea

## What stayed with us

These decentralized inference projects make the latency cost of sharding a model across machines tangible. The comparison helps explain why a single-Mac specialist is a different system.

02 · The local translation

## What we did with it

Together these projects illustrate two shapes of decentralized inference: sharding work across machines and making complete models available across a network. The distinction sharpens our understanding of per-token latency and what locality buys a small specialist.

03 · The boundary

## Where the comparison stops

We have not installed or benchmarked either system for PostTrainLLM. Their networked design is a comparison at the distributed boundary, not a substitute for a verified local decode and task-completion gate.

Source trail

## Follow the work

- Original project [teale ↗](https://teale.com/)
- Further reading [Petals source ↗](https://github.com/bigscience-workshop/petals)
- Our context [Our Mac mastery map ↗](https://posttrainllm.com/docs/learn/mac-mastery-map/)

Independent appreciation. The named projects have not endorsed or affiliated with PostTrainLLM.