On Keryx, inference is ON-CHAIN. That is not a slogan: the request is a transaction paid in
$KRX, the miner runs the model on their GPU, the answer is written into a block, and the chain itself pays the miner in
$KRX. As far as I know, we are the only ones doing this. Every "AI + crypto" project I have looked at does the same thing: the token is there to vote or to speculate, and the actual inference is served off-chain by a regular provider (Pearl
@prlnet and TSC
@Tensorcash both do OFF-CHAIN inference; Pearl, for example, goes through
@togethercompute). With us, the miner IS the server. The mining computation itself is an execution of the model (Proof of Model): no model in VRAM, no block.
Where the coin comes in: it is the only way to buy an inference. No account, no API key, no credit card, no KYC, no filter: you send
$KRX, the network answers you. Today 5 models are served (GLM-4 9B, Gemma-4 12B, Qwen3.5 9B, Kimi-Linear 48B, Qwen3.6 27B), and I am currently working on integrating much bigger models (DeepSeek V4, Kimi K3, GLM 5.3). The same GPU that secures the chain gets paid twice: for the block and for the answer.
Why it is huge: autonomous AI agents are arriving everywhere, on
@ethereum , on
@solana , in dApps. An on-chain agent cannot open an
@OpenAI or
@AnthropicAI account or pull out a
@Visa. It needs a source of intelligence it can pay for with what it has: a token. That is exactly what Keryx sells, and that is phase 6 of the roadmap: the bridge so those agents can call Keryx directly. Every request consumes
$KRX and 100% of fees are burned. Demand for the token is mechanically tied to AI usage, not to a vote.
Yes, demand is still small: the market has not yet understood what we have built. That is precisely what being early means.