What is being demonstrated here?
Three cyclonechain nodes are running on the network, each with
its own VM wrapper running llama‑3‑8b (all model layers).
The process is as follows:
- Consensus (≈ 7 seconds).
The nodes reach a common agreement on two values:
- The input prompt embedding
- A decentralized random seed for sampling
- Deterministic inference.
Once the embedding and seed are fixed, the behavior of
llama‑3‑8bbecomes deterministic. Each node autonomously runs a full pass of the model inside its VM and writes the response to a local VM log. - Optimistic result matching. Since the computations are deterministic, the responses from all nodes should match; no additional consensus rounds are required. If proof is needed in the future, selective verification of response hashes can be added.
Response Time. Consensus consistently takes ≈ 7 seconds. The VM processing itself depends on the load of each node and can take from tens of seconds to several minutes (we are running the nodes on minimal resources—this is a POC, after all).
Why is it so slow, and what's next?
Currently, we are demonstrating a full inference, but the ultimate goal is
to compute the results of one or more model layers and
aggregate them across the nodes. This approach will reduce the load,
allow for parallel computation, and significantly decrease latency.
Full design breakdown
here
The block history can be viewed here. If you want to see how the same setup works with StableDiffusion 1.5, contact us—we only enable GPU hosting for demonstrations.
Go to the Llama test page →How to run a query?
-
Install the cyclone wallet extension (Chrome / Chromium):
chrome web store → cyclone‑wallet -
In the wallet, add the custom
cycloneLLMnetwork and enter one of the addresses in the “Gateway / RPC” field:http://37.60.225.104:9231
or the address of your local node (if you have already deployed it).1
2
3
4
- Open the Test page using the button above, allow the wallet connection, and enter your prompt. The result of the consensus on the Embedding, random-seed, and LLM call ID will appear after 8‑12 seconds — this is the time for consensus and writing to the blockchain. The inference result will appear after the request is processed by the VM of the node connected to your wallet.