Question 1

What is a decentralized LLM?

Accepted Answer

A decentralized LLM is a large language model whose inference runs across a network of independent compute providers rather than inside a single company's data center. With LibertAI, open-weight models such as Qwen3.6-35B-A3B, Gemma 4 31B, and DeepSeek V4 Flash are served on Aleph Cloud, a decentralized cloud network. No single operator owns the compute or can unilaterally read data or shut the service off.

Question 2

How does decentralized AI inference work?

Accepted Answer

You send a request to a standard OpenAI-compatible endpoint at https://api.libertai.io/v1. The Aleph Cloud network schedules that request onto an available node operated by an independent participant, the model runs the inference there, and the response comes back. Sensitive workloads can be routed to nodes that run inference inside a Trusted Execution Environment, which you can verify with remote attestation. Node operators are paid through Aleph Cloud's resource token economics, which keeps capacity online without a central operator.

Question 3

Is decentralized inference slower?

Accepted Answer

For most workloads, latency is comparable to centralized providers. Inference still runs on real GPUs and CPUs; the difference is who operates them, not the underlying hardware speed. A request is scheduled onto a capable node and runs there. The decentralization affects governance and trust, not the per-token compute time in any fundamental way.

Question 4

What is Aleph Cloud?

Accepted Answer

Aleph Cloud (aleph.cloud) is a decentralized cloud network. Independent node operators provide compute and storage, coordinated by the network rather than by one company. Its resource token economics incentivize operators to keep nodes online. LibertAI runs its model inference on this network, which is what makes the inference decentralized rather than hosted by a single provider.

Question 5

Can decentralized AI be censored or shut down?

Accepted Answer

There is no single operator who can read the data, filter answers, or flip a central kill-switch, because no one company owns the compute or controls the coordination. That is what censorship resistance means here: it is a property of the architecture. The models are also open-weight and the serving stack is open source, so even in the worst case you can self-host the same models yourself on your own hardware.