Run unified-memory inference on OsxNode with a dedicated physical Mac. Before you start, check who can connect, what happens to your files when the rental ends, and how to request compensation if service is interrupted.
OsxNode M4-16: M4 / 16GB unified memory / 256GB SSD. A dedicated machine does not include automatic backups. Export your models, datasets and experiment results yourself.
Each Mac mini M4 is allocated to one tenant for the duration of the rental. Isolation applies to the entire physical machine, not just a desktop session.
Your machine's unified memory and GPU cores aren't shared with other tenants. When loading a model, you can assess resource use based on the machine's own memory pressure and running processes, without accounting for capacity reserved for someone else.
Keep in mind: a dedicated machine doesn't increase capacity. Model weights, context cache, inference frameworks and system overhead all count toward the 16GB unified memory limit.
Local storage isn't shared with other tenants. OsxNode doesn't access files on your machine; you're responsible for managing downloaded weights, datasets, environment files and output directories.
Keep in mind: dedicated storage is not an off-site copy, and it doesn't mean files will remain available after your rental ends. Back up important results to a location you control before then.
Use the dashboard to manage orders and nodes; use your remote desktop password and SSH key to access the machine. Having one set of credentials is no reason to share another.
Sign in with your own account and check the order, node name and rental period before connecting. Protect your account email and sign-in credentials. Don't forward verification codes or sign-in details, or share full-page screenshots that contain sensitive information.
Sign out after using a shared computer. To report a delivery issue, describe it in a support ticket for the relevant order in the dashboard.
Get your connection details from the dashboard and change the initial password as soon as you sign in. Store your machine password separately from your dashboard credentials. Don't put it in a project repository, public notes or inference logs.
See the remote access guide for browser-based remote desktop, VNC and screen-sharing instructions. Public pages don't provide a direct link to an instance desktop.
If you need SSH, add your public key to the rental machine and keep the private key on a trusted device. Use separate keys for different purposes and review authorized keys regularly; remove any you no longer use.
Access tokens and environment variables in automation scripts are credentials too. Check scripts for passwords before backing them up, and revoke any external access you no longer need when your experiment is finished.
If you don't renew, the machine shuts down when your rental expires. Shutdown, disk erasure and OS reinstallation are separate steps. A stopped service doesn't mean your files are backed up or that you can still sign in.
Stop tasks that are writing results, then export model deltas, checkpoints, processed data and environment dependency lists. After transferring large files, check their size or checksum to make sure the copies are readable.
Keep in mind: starting a download doesn't mean a backup is complete. Allow time for transfer and verification; don't leave your only copy until just before expiry.
Nodes that aren't renewed shut down automatically at expiry, with no further charges. Don't rely on inference, training or build tasks continuing after shutdown.
Keep in mind: we don't promise a fixed data-retention period or guarantee that you can regain access after shutdown. Renew before expiry if you need to keep using the machine.
Once a rental machine is no longer in use and enters the reclaim process, its disk is erased and the OS reinstalled. The machine is cleaned before it's handed over again; the previous tenant's environment isn't passed directly to the next tenant.
Keep in mind: we don't guarantee file recovery after erasure. This process doesn't replace your own backups. Check the status of a specific order through a support ticket.
Nodes are intended to run year-round, 365 days a year. The commitment below is a service uptime target, not a record of past performance or a guarantee that any individual inference job will complete.
Equivalent to about 43 minutes of downtime over 30 days.
Calculation: 30 × 24 × 60 × 0.1% = 43.2 minutes. This illustrates the commitment; it doesn't represent actual downtime.
Downtime beyond the allowance is converted to account credit in proportion to the service fee for the relevant period. Submit a request through a dashboard support ticket during that period, and include the node name and downtime window.
Record the start and recovery times, including the time zone, and attach connection errors or other details we can verify. Don't include machine passwords, SSH private keys or full payment card details in attachments.
Force majeure and issues caused by your own actions are excluded. Insufficient memory for a model, a user process exiting or a problem with your local network connection doesn't by itself count as node downtime; we'll review the records to determine the cause.
If the interruption is ongoing, submit the times and details you have during the current service period, then add the recovery information to the same ticket once service is restored.
Our nodes are currently located in Singapore, Tokyo, Seoul and Hong Kong. Choosing a node means choosing where its physical machine is located; it doesn't automatically copy your models, results or running tasks to another location.
| Physical node | Region served | Deployment & recovery |
|---|---|---|
| Singapore | Southeast Asia | Physical machines are provisioned separately; you arrange any cross-region copies. |
| Tokyo | East Asia | Physical machines are provisioned separately; you arrange any cross-region copies. |
| Seoul | Northeast Asia | Physical machines are provisioned separately; you arrange any cross-region copies. |
| Hong Kong | South China and Southeast Asia | Physical machines are provisioned separately; you arrange any cross-region copies. |
Power and upstream connectivity are part of the physical node infrastructure. The number of locations doesn't tell you how many power feeds or carriers an individual machine has, or whether automatic failover is available. We don't claim data center certifications that haven't been provided.
If your procurement requirements call for dual power feeds, a specific upstream provider or a redundant topology, check the details for the specific node by email. This page doesn't confirm equipment-level specifications.
Keep dependency versions, startup parameters and result-export steps somewhere you control. To restore an experiment, use your records to rebuild the environment instead of relying on temporary actions in a remote desktop as your only documentation.
If you need to switch nodes, plan for file transfers and environment setup. Multiple locations don't provide automatic migration of in-memory state, task handoff or data synchronization.
You need the appropriate rights to install software, download model weights and upload data. Dedicated resources don't change the terms of model licenses, data permissions or third-party services.
Review model weight licenses, dataset sources and whether processing on a third-party hosted device is allowed. If you're working with personal information, partner data or internal code, secure the required permissions before transferring it to your rental machine.
Manage the software you install, the services you expose and their access permissions. If an inference service needs to be accessible over a network, configure authentication and limit who can reach it. Don't expose model APIs, file directories or debug ports publicly.
Back up your experiment results and reproducible configurations in advance, and revoke tokens and permissions you no longer use. You must not use the machine for cryptocurrency mining, bulk unsolicited email, attacks on others or any other activity that violates the Terms of Service.
Choose a rental period that fits your experiment, and prepare your environment and export plan. If you need to confirm data requirements or node specifications, contact support before uploading files.