Compare the actual offer against your workload: supported models, deployment control, trust, throughput, availability and total cost. Product names or a headline hourly price do not establish equivalent service.
| Requirement | Current Mac Compute behavior |
|---|---|
| Deployment | One approved Apple Silicon Mac allocated to one customer for 30 days, with managed text inference. |
| Models | The pinned Qwen3.8 27B catalog in 4-bit, 8-bit and BF16. Check each model's memory/context limits and current stock. |
| API | Text Chat Completions with optional SSE streaming. Scoped renter keys; supported OpenAI client base-URL integration. |
| Limits | One running job per Mac, up to 4,096 output tokens, model context limits, input-byte limits and request/deadline controls. No per-token billing. |
| Price | Admin-approved USD quote per node for 30 days, paid from account balance. Manual renewal; review current terms each time. |
| Control | Choose a catalog model fitting your tier. No SSH, customer shell, arbitrary model import, training, fine-tuning or customer containers. |
| Trust | The supplier owns/administers the Mac. This is not confidential computing or zero retention. |
| Availability | Only currently eligible free machines are rentable. Recorded outages pause remaining rental time. Early access has no SLA. |
Check current eligible machines before adding funds for Compute.
Current catalog · Default prices · Effective node quotes and stock
| Area | Verify in the selected offer |
|---|---|
| Model fit | Exact model revision, quantization, context, output cap and required quality. Can your application use that API without unsupported features? |
| Deployment responsibility | Do you receive an inference endpoint, a job runner, a container or a machine? Who installs, updates, secures and restarts the serving runtime? |
| Access and isolation | Who administers the host? Which resources are exclusive? Who can access prompts, logs, operating-system memory and backups? |
| Performance | Measure first-token latency, completed-response latency and usable throughput with your own prompt lengths and concurrency. A chip name is not a workload benchmark. |
| Commercial terms | Review the current quote, minimum commitment, idle charges, storage/network charges, renewal and cancellation terms, and supported payment methods. |
| Reliability | Check current supply, recovery behavior, maintenance, support response and any contractual SLA. A listed machine is not proof of a completed customer workload. |
For Nosana, io.net and Akash, read the selected offer and its current provider documentation directly. This checklist does not assert their prices, payment-token requirements, deployment isolation or supported features.
Mac Compute is a candidate when its supported text model, managed API and 30-day commitment match your application. Check current rentable capacity and validate your own quality, latency and concurrency requirements before expanding.
If you need CUDA, fine-tuning, arbitrary models, tools, structured response format or multi-machine deployment, identify an offer explicitly supporting those requirements. They are not capabilities of the current Mac Compute service.
Compare total committed cost against useful completed work under the same test conditions. Include idle time, failed requests, setup effort and your operational responsibilities. A flat 30-day price and an hourly infrastructure quote are different units; neither establishes which is cheaper for your application.
Read the complete Mac Compute guide API and operating reference