Lend your machine to the Korai network. Get a majority share of what users pay, per second of compute served.
Ein neuer Host bedient erst Anfragen, wenn ein Administrator das Konto freigegeben hat; bis dahin versucht dein Worker von selbst, sich erneut zu verbinden.
Paid only when your machine serves a real inference job. No reservation fee, no idle charge.
The fleet is audited every 30 days and updates itself on your machine, with no service interruption and no action from you.
Requests reach your machine without the requester's account, email or IP address. You do see the text your machine processes, so host only if you are comfortable with that.
M1/M2/M3/M4: läuft direkt ab Werk. Je mehr Unified Memory du hast, desto fortgeschrittener das Modell, das Korai lädt — bis hin zu mehreren gleichzeitig geladenen Modellen bei den großzügigsten Konfigurationen.
RTX 30/40/50, A-Serie, H-Serie. Das Skript wählt automatisch das fortgeschrittenste Modell, das dein VRAM laden kann, mit einem optimierten NVFP4-Format auf Blackwell-Karten (RTX 50).
The full cycle on the host side, no gray areas: what the script does, what your machine sees, what ends up in your bank account.
One command to copy-paste into the terminal (macOS, Linux, Windows). The script detects your hardware and installs the matching inference runtime.
Your browser probes your GPU via WebGPU to suggest the right profile (Apple Silicon, NVIDIA, AMD). Nothing leaves your machine before you explicitly launch the worker.
Your worker downloads the model(s) from Korai's signed manifest based on your available VRAM. The fleet is audited every 30 days: when a model changes, the new one downloads in the background and swaps in with zero downtime (hot-swap), with no action from you.
The orchestrator sends requests to your GPU over a TLS connection. Your worker and the model engine read them in clear, and so can you: hosts see the text they process.
Every inference served triggers an Ed25519-signed cryptographic receipt. 75% of what the user paid comes back to you, computed on the actual second of compute delivered, never on a reservation.
Connected machines, jobs served, cumulative earnings: all visible at /platform/machines. Aggregated SEPA payout on the 5th of the following month.
An order of magnitude, not a promise: actual earnings depend on network demand while your machine is online.
75% of what the user pays comes back to you directly (25% to Korai), roughly €0.375 per million tokens generated, paid only for compute actually served. No reservation fee, no minimum hours.
| Hardware | Indicative throughput | Estimated gross monthly* |
|---|---|---|
| RTX 4090 | ~196 tok/s (~30B Q4 model) | ~€19-86 |
| RTX 5090 | ~850 tok/s | ~€83-372 |
| Apple Silicon Mac, 32-64 GB (M2/M3 Max) | ~120 tok/s | ~€12-52 |
* Range based on 2.4 to 10.8 h/day of effective availability (10% to 45% of the time, depending on network adoption phase), at the host payout rate of €0.375/M output tokens. Gross amounts, before electricity. RTX 4090 throughput: docs/MODEL_FLEET.md. Methodology and payout split: docs/PRICING.md.
Teile die Leerlaufzeit deines Gaming-PCs oder Mac Studios und werde für jede Berechnungssekunde bezahlt. Du behältst die Kontrolle: leg deine eigenen Zeiten fest, höre auf, wann du willst.
Keine Kreditkarte. Du kannst den Worker jederzeit deinstallieren.