Your Mac can run low on memory before the model even finishes loading. OsxNode turns access to a dedicated Mac mini M4 into a service you can rent by the day—not a decision to buy a computer.
For individual developers and researchers using MLX for inference or small-scale fine-tuning. First check whether your workload fits in 16GB of unified memory, then choose a rental term for your experiment.
From $21.6/day. Pay upfront by the day, week, month or quarter. Your Mac shuts down at the end of your term unless you renew, with no further charges.
OsxNode · Service scope
One physical Mac, assigned to one tenant for the duration of the rental term.
The CPU and GPU share the same memory pool. Extra SSD storage gives you more file space, not more unified memory.
You may need to run one model, a set of experiments or a single build—not buy equipment for the next few years.
When trying MLX locally, the first challenge is often not the command itself but memory shared by model weights, context cache and other apps. Buying a new computer to test one idea may not make sense for the project or its timeline. And if you only need a weekend, you may not want to commit to a long-term plan.
OsxNode Cloud Mac was built for this specific use case: the machine stays whole, while you choose a term by the day, week, month or quarter. Set up your environment, weights and workload during your rental. When the experiment is done, export your results; your Mac shuts down when the term ends unless you renew.
This isn’t a single machine divided into quotas for multiple users. We divide rental time, not unified memory or GPU cores. You still need to check whether your model fits the available configuration and allow time to upload data, install dependencies and back up results.
There’s currently one model available. Your choices are based on project duration, location and optional storage—not hidden hardware tiers.
OsxNode M4-16: M4 chip, 16GB unified memory and a 256GB SSD. The physical Mac is dedicated to you for your rental term, with full access to the macOS desktop and command line.
Limit: No models with more than 16GB of memory are currently available. Storage upgrades do not add memory.
Singapore, Tokyo, Seoul and Hong Kong are currently available. Choose based on your network connection and file transfer needs; location alone doesn’t guarantee task speed.
Limit: Check the console for real-time availability. Remote desktop responsiveness also depends on your local network.
Day: $21.6; week: $58.2; month: $107.8; quarter: $293.2. Your Mac is provisioned on the day your payment is confirmed. Delivery is complete when the console shows your location and connection credentials.
Limit: Provisioning on the same day doesn’t mean instant delivery after payment. Your Mac shuts down when the term ends unless you renew, so plan your work and backups accordingly.
Payment is accepted only via USDT-TRC20 or Visa, Mastercard and Amex through Stripe. All payments are settled in USD. Available payment gateways are confirmed by the backend.
Infrastructure delivery and model workloads are separate things. Clear service boundaries help you decide whether moving your project here is the right fit.
We don’t split a physical Mac’s unified memory or GPU cores among multiple tenants. Dedicated means the hardware is assigned to you for your rental term—not that your workload can exceed its physical capacity.
Each machine is assigned to one tenant for each rental term. Check the console for real-time availability; listed options and current availability are confirmed separately. The console status—not an assumption—is the source of truth.
The current configuration has up to 16GB of unified memory and no additional GPU accelerator. We don’t recommend running 13B or larger 8-bit models on this configuration. Longer contexts require even more memory.
You’re responsible for choosing models, downloading weights, installing dependencies, tuning parameters and checking outputs. We don’t install model weights or guarantee a particular model’s generation speed, training convergence or experiment results.
Choose where the machine is located. Your console provides instance details and connection credentials. A location name doesn’t indicate a different hardware tier.
The service currently covers four locations: Singapore, Tokyo, Seoul and Hong Kong. After provisioning, check the location and order in your console before connecting to your assigned machine.
We don’t promise certifications, measured bandwidth or power redundancy details that haven’t been provided. A machine’s location alone doesn’t guarantee network quality.
After logging in, find your connection details in the instance view and open the remote desktop in your browser. For command-line access or file transfers, you can also configure SSH or VNC using the connection instructions.
Change the initial password when you first connect, and keep your keys secure. Use the connection details shown for your provisioned instance.
Short-term rentals work best for tasks with reproducible environments, predictable resource needs and results you can export before your term ends.
Start by checking that a quantized model loads, then gradually increase context or batch size. Track weight size, peak memory and output rather than judging feasibility by parameter count alone.
Small-scale LoRA fine-tuning also requires you to assess training state, activations and sequence length. A model that can run inference won’t necessarily fit for fine-tuning.
Before you start: Test with a smaller sample and shorter sequences. If memory pressure rises, reduce the workload; don’t treat swap as extra memory.
Move your source code, dependency manifest and build commands together. Restore dependencies first, then run a full build. Save caches separately from build artifacts to reduce the data you need to export at the end.
Model inference and compilation compete for the same 16GB memory pool. If you need both, schedule them at different times instead of assuming they can run in parallel.
Before you start: Check required development tool versions, signing materials and disk space. You’re responsible for providing project dependencies and licenses.
Set up your own self-hosted runner on the machine and script its installation and working directory. Use a clear job queue to schedule builds and upload results during your rental term.
A dropped remote connection shouldn’t be your task management strategy. Use tmux to keep long-running jobs alive, then check the logs and exit status when they finish.
Before your term ends: Remove the runner registration, export the artifacts you need and delete credentials. Once your rental ends and the Mac shuts down, you can’t rely on it to run jobs.
Hardware assignment, memory capacity and billing are separate commitments. Don’t mistake one for another.
The plans page lists four rental terms and optional storage. The help center covers connections, environment setup and pre-expiry checks. Confirm your workload fits in 16GB of unified memory before renting.