How to mine
Cortex rewards work that can be checked: a better model proven in a duel, or a real bug an operator reproduces. Bounty is the live challenge and takes signed submits through the public gateway. Hypertrain is a preview. relearn, relearn-image, relearn-agent, relearn-mm, design, prism, proof, opentype are off: they have no trust-root row, so they earn nothing.
Before spending time on Bounty: a report only pays after an operator reproduces it and prices its severity, and scoring can be paused. Check the status card on the Bounty page first. A ready status is not an end-to-end payment guarantee.
Get the miner app
Copies a single command to your clipboard. Paste it in a terminal on your own machine and the miner app installs itself.
Point your LLM at the network guide. It installs the app, checks what is scoring today and walks you through the first submission.
https://network.cortex.foundation/llms.txtThe app talks to network.cortex.foundation by default. Prefer an agent? Hand it the network guide instead.
Challenges
Hypertrain preview
Hypertrain trains one shared model across Subnet 100 miners with decentralized training. Each miner trains locally for many inner steps, then the group syncs pseudo-gradients once per outer round through an outer Nesterov optimizer, so it needs a fraction of the bandwidth of DDP. It is not live: the page shows simulated data, labelled as such, and there is nothing to submit yet.
OPEN THE HYPERTRAIN PREVIEW →What every challenge pays for
Bounty pays precision times severity, and the triage-noise ratio stays off the visible score. Missing evidence fails closed: a host that cannot score refuses instead of inventing a verdict. Hypertrain pays nothing while it is a preview. Check the Status page before you spend anything.
Keys you never share
Never put mnemonics or signing keys in a miner client or an agent. Keep pairing secrets in your environment, never in a repo. Changes to the network software go through operator review; that is not a miner step.