Share Compute
Your Mac serves the models. You earn API credit.
Share Compute connects an idle Apple Silicon Mac to the QuickSilver Pro pool. It answers real customer requests for the open-weight Qwen 3.8 27B with Rapid-MLX, and every token it serves earns credit you can spend on any model we host.
80% of what customers pay, as API credit. No port forwarding. Stop anytime.
- 80% of what customers pay, as API credit.
- No port forwarding.
- Stop anytime.
API credit, not cash.No cash-out, bank transfer, or crypto payout, and credit cannot be transferred.
See what your Mac could earn
Try a 30-day estimate before connecting your Mac.
Enter the tok/s Rapid-MLX shows on your Mac. The speed-only presets are scenarios, not machine benchmarks.
This is an assumption: the share of connected time spent serving requests depends on pool traffic.
Estimated API credit / month
$3.95$3.95
Estimated output tokens / month: 3,456,000Estimated output tokens / month: 3,456,000
API credit, not cash · settles hourly
That credit buys ≈
7.9 million output tokens of GPT-6 Luna.7.9 million output tokens of GPT-6 Luna.
Estimate: tok/s × hours/day × 3,600 × 30 days × serving share. Credit uses the catalog output price × 80%, up to the cap. Actual payouts follow what the customer paid.
This estimate counts output tokens only. Real earnings also include input tokens. The buying-power estimate excludes input costs.
How this estimate is calculated
Estimate: tok/s × hours/day × 3,600 × 30 days × serving share. Credit uses the catalog output price × 80%, up to the cap. Actual payouts follow what the customer paid.
This estimate counts output tokens only. Real earnings also include input tokens. The buying-power estimate excludes input costs.
The pool is new and traffic is limited. Think of this as offsetting your API bill, not income. These estimates are assumptions, not promised earnings.
How it works
Eligible requests are routed to a pooled Mac. When no Mac is free, a cloud provider answers instead, so customers never wait on you.
Your Mac dials out to the relay over one encrypted WebSocket. Nothing ever connects in to your network.
01 · Pool relay ↔ Your Mac
One outbound connection
No inbound ports, no port forwarding, no public IP. Your Mac opens a WebSocket to the relay, and requests arrive on it in real time. There is no job queue to poll.
02 · Your Mac
Real requests, served locally
Rapid-MLX runs the model on your Mac's GPU and streams answers back. The pool routes each conversation to the same machine, so your Mac's prompt cache keeps follow-up turns fast.
03 · + API credit · cloud fallback
Credit for every token
QuickSilver counts the tokens itself and credits your account balance. If your Mac is busy or offline, the request goes to a cloud provider instead, and the customer never notices.
If the Mac is busy or offline, the API sends the request to a cloud provider. No credit settles for that request.
Get started
Start sharing in four steps
One key, one install, one command. Each step opens as you reach it — or jump to any step.One key, one install, one command. Open one step at a time.
Step 1
Create a provider key
Sign in and open Share Compute in your dashboard. Create a provider key (it starts with qsppk-). It can only register machines for your account. It cannot spend your balance.
Open the dashboard →Step 2
Install Rapid-MLX
Rapid-MLX is the open-source MLX inference engine for Apple Silicon. Requires 0.17.0 or later.
curl -fsSL https://rapidmlx.com/install.sh | bash
export PATH="$HOME/.local/bin:$PATH" # so this terminal finds rapid-mlx
# or: brew install rapid-mlxStep 3
Start sharing a model
Run this in Terminal and paste your provider key when prompted, so the key stays out of your shell history. Use a different --worker name on each Mac, and press Ctrl-C to stop.
rapid-mlx share qwen3.8-27b --quicksilver --worker "$(hostname -s)"The first run downloads about 15 GB and may take a few minutes. Once ready, the terminal shows “serving qwen3.8-27b to the QuickSilver pool”.
Certificate error on the first run?
Certificate error on the first run after a Homebrew install? Run brew postinstall openssl@4, or use the installer above.
Disk caches, and cleanup if you shared before 2026-10-09
Pool mode disables Rapid-MLX’s prompt disk caches automatically — there’s nothing to set.
If you shared before 2026-10-09, delete old caches once: rm -rf ~/.cache/rapid-mlx/prefix_cache.
Prefer an app? Use Rapid-MLX Desktop
Prefer an app? In Rapid-MLX Desktop, turn on Share Compute under Settings → Experimental, pick a model in Live Pool, and click Connect & serve.
Download Rapid-MLX Desktop ↗Step 4
Watch your credit arrive
In the dashboard, “Your nodes” shows whether your Mac is online and what it earned this month. Earnings are updated hourly.
Open the dashboard →To see earnings in Rapid-MLX Desktop instead, add the same provider key under Credits. It can only read your earnings there — never spend your balance.
What your Mac needs
- Recommended
- Any Apple Silicon Mac
- Unified memory
- 32 GB+
- Concurrency
- up to 2 requests
- Network
- outbound only
Any Apple Silicon Mac with 32 GB+ unified memory. 48 GB+ uses the TensorFold runtime for the best throughput; 32–47 GB automatically uses the lighter MTP runtime. Below 32 GB, this model can’t be shared yet. By default a node handles up to two requests at a time.
Not on a Mac?
The pool protocol works with any machine that runs an OpenAI-compatible server (vLLM, llama.cpp, Ollama, LM Studio). Use the open-source reference client to connect it.
quicksilver-pool-node on GitHub ↗Read the full Rapid-MLX Share Compute guide ↗Earnings
What you earn, exactly
See how one $1.00 request becomes API credit.
Example · one $1.00 request
Tracks list price
Follows our price list
Your rate is not locked in. It tracks QuickSilver's price list automatically, so when a model's price goes up or down, your earnings change with it.
80%
$0.80 reaches your API balance
On this $1.00 example, QuickSilver keeps $0.20 and your balance receives $0.80.
Hourly
Settled hourly
Work is credited to your balance every hour. A delayed settlement is paid later and never paid twice.
$20 / node / mo
Monthly cap per node
To limit fraud, each node can earn up to $20 per calendar month (UTC). Anything above the cap is not paid and does not carry over.
Credit only
API credit, not cash
Earnings go to your QuickSilver account balance and pay for API calls on any of our models. There is no cash-out, bank transfer, or crypto payout, and credit cannot be transferred.
$0
Your own traffic doesn't count
API calls paid for by your own account earn nothing, even when your Mac serves them. You can't pay yourself.
How the 80% works
You earn 80% of what customers pay for every request your Mac serves. Your rate follows QuickSilver's price list automatically: when a model's price changes, your earnings change with it. QuickSilver keeps 20%, and a payout can never exceed what the customer paid. Earnings are paid hourly as QuickSilver API credit, not cash, up to $20 per node per month.
80% of what customers pay
For every request your Mac serves, you earn 80% of what the customer actually paid for it. QuickSilver keeps 20%, and a payout can never exceed the customer's charge.
The model you can serve today
Qwen 3.8 27B is the pool's launch model; customer requests use qwen3.8-27b-decentered. Earnings below use that model's 30% lower community price (80% of the QuickSilver price, per 1M tokens); each request actually pays 80% of what that customer was charged. Cloud fallback calls earn no contributor credit.
| Model | Recommended Mac | You earn (input / output per 1M) |
|---|---|---|
| Qwen3.8 27Bqwen3.8-27b | Any Apple Silicon Mac with 32 GB+ unified memory. 48 GB+ uses the TensorFold runtime for the best throughput; 32–47 GB automatically uses the lighter MTP runtime. Below 32 GB, this model can’t be shared yet. | $0.1904 / $1.1424 |
Your rate follows our price list automatically. More models will be added later.
Read the contributor terms →Your responsibilities as a contributor
Your Mac processes real customer prompts and completions. Customers trust us with that data, so by sharing compute you agree to handle it with care.
You must: Don't log, keep, or read customer content.
Run the node with request logging off, and never store, inspect, or reuse prompts or completions. Pool mode disables Rapid-MLX’s prompt disk caches automatically — there’s nothing to set.
QuickSilver will: Usage is measured by us.
QuickSilver counts tokens itself and ignores what the node reports. Nodes whose reports don't match are flagged for review.
You must: Don't tamper with outputs.
Serve the listed model unmodified, and never alter, filter, or fabricate responses.
QuickSilver will: We can remove nodes.
We may remove a node, and withhold credit earned through abuse, if it breaks these rules or degrades service. To remove your machine permanently, open “Your nodes” → Remove in the dashboard. If you need help, email hello@quicksilverpro.io.
You must
Run the node with request logging off, and never store, inspect, or reuse prompts or completions. Pool mode disables Rapid-MLX’s prompt disk caches automatically — there’s nothing to set.
QuickSilver will
QuickSilver counts tokens itself and ignores what the node reports. Nodes whose reports don't match are flagged for review.
You must
Serve the listed model unmodified, and never alter, filter, or fabricate responses.
QuickSilver will
We may remove a node, and withhold credit earned through abuse, if it breaks these rules or degrades service. To remove your machine permanently, open “Your nodes” → Remove in the dashboard. If you need help, email hello@quicksilverpro.io.
Contributor terms are part of our Terms of Service·Privacy Policy→
Frequently asked questions
It depends on how much traffic the pool routes to your model and how fast your Mac serves it. You earn 80% of what customers pay for every request you serve, up to $20 per node per month. Today the pool is small, so think of this as offsetting your own API bill, not as income.
No. Earnings are QuickSilver API credit. It pays for API calls on any of our models, but it cannot be withdrawn, transferred, or converted to cash or crypto.
No. Your Mac makes one outbound, encrypted WebSocket connection to our relay, and requests arrive over it. Nothing connects in to your network.
While it serves a request, the model uses your GPU and memory. By default a node handles up to two requests at a time. You can stop sharing whenever you want to use your Mac: press Ctrl-C, or click Stop sharing in the desktop app.
Nothing breaks. The pool stops sending you requests and serves customers from cloud providers until your Mac reconnects. Rapid-MLX reconnects automatically.
You must not. Prompts and completions pass through your Mac to be processed, but contributors agree not to log, store, or inspect them. Keep request logging off. Pool mode disables Rapid-MLX’s prompt disk caches automatically — there’s nothing to set.
To stop anyone from paying themselves, calls billed to the node owner's own account are excluded from payouts.
Stop sharing (Ctrl-C or Stop sharing) and your Mac stops receiving work right away. To remove its registration permanently, open “Your nodes” → Remove in the dashboard; earned credit is kept. If you need help, email hello@quicksilverpro.io.
Put your Mac to work
Create a provider key, run one command, and your Mac starts earning API credit on real traffic.
The pool is new and traffic is limited. Spend the credit on any model we host.