Use Kunagi
A chat app and API over many models. Send a prompt, the router picks eligible GPU supply, and tokens stream back in real time.
Open the appDecentralised
GPU inference.
Kunagi Systems is a privacy-first compute network. Send a request, the router picks eligible GPU supply, and tokens stream back.
In partnership withIBM
App · APIRouter
GPU workerToken stream
8-way
GPU fan-out per request
Streamed
Token-by-token response
Permissionless
Any GPU can join
Usage-priced
Pay per token, not per hour
How a request gets routed.
The router selects eligible supply by model, availability, and measured speed. The worker executes and tokens stream straight back.
Worker benchmarkBlackwell-class reference
kunagi-open-8bfirst token 0.24s
168 tok/s
kunagi-mistralfirst token 0.21s
182 tok/s
kunagi-reason-r1first token 0.38s
96 tok/s
kunagi-reason-minifirst token 0.14s
340 tok/s
Every worker runs kunagi worker benchmark on join and reports its own numbers. The router weights routes by measured tokens per second, not by advertised hardware.
NVIDIA · Blackwell
GB202 / GeForce RTX 5090
21,760
CUDA cores
6805th gen
Tensor cores
1704th gen
RT cores
32 GBGDDR7
Memory
512-bit
Bus width
92.2B
Transistors
4 nmTSMC 4N
Process
575 W
TGP
Memory bandwidth by generation
RTX 3090
936 GB/s
RTX 4090
1008 GB/s
RTX 5090
1792 GB/s
Specs from NVIDIA's RTX Blackwell architecture whitepaper. Kunagi workers report the same fields, VRAM, cores, bandwidth, for routing.
Usage is counted, earnings are credited, and prompt context is discarded once the response completes.
Two sides of the network.
Provide compute
Run the worker client, benchmark your machine, list supported models, and accept routed jobs. Earnings are credited per usage.
Become a providerStake and earn
Token holders stake in a vault for a share of rewards. Workers post a bond for higher tiers. The bond is slashed if verification fails.
Read tokenomics