The AI of the People

Rate limits are per API key: 20 requests/min on the free models, 60/min on the paid ones at pay-per-token. A subscription raises the paid limit — see plans. Paying with credits, one credit buys one token on a 70B/72B model; the smaller models cost 1.15 credits per token, because serving them on the same hardware is not proportionally cheaper.

These models are not answering requests today. No GPU is attached to the serving side, so /v1/models and /v1/chat/completions return 503. This is the catalogue we intend to serve, published so you can price against it — not a menu you can order from yet.

Inference is offline — there is no GPU on the serving side and API calls return 503. There is also no checkout: subscriptions cannot be purchased, so the plans below have no buy button and we are not taking payment or card details for any of them. These are the rates we will charge when serving comes up, published now so the terms are on the record before there is anything to sell.

The Free plan covers the free and standard models; every paid plan covers free, standard and premium. Enterprise includes custom contracts and dedicated support. Credits shown are per month and roll over for up to two cycles. Rate limits are per key, per minute.

Where we are cheaper: the 70B class. At $0.19 we are below Groq's $0.59, Together's $0.88 and Fireworks' $0.90 — roughly a fifth of the most expensive of the three.
Where we are not: the 8B class. At $0.22 we are dearer than all three, Groq's $0.10 included. We serve small models for completeness and for miners spending credits they earned, not to win on price.

Z.ai is on our comparison sheet too, though the models are not the same ones and so do not line up in the table above: GLM-5 at $1.43, GLM-4.7 at $0.50, GLM-4-Flash free.

Competitor prices recorded 2026-08-01 and not re-checked since; verify against each vendor's own pricing page before you rely on them. A price we cannot yet honour is not a saving — this comparison is here so the intended economics are legible, not to claim you are saving money today.

Before you register, know what you are getting. Mining works today and every accepted share is recorded against your address. Inference capacity is offline — there is no GPU attached — so the credits you earn cannot be spent yet and API calls return 503. You are registering to start accruing, not to start calling. We would rather you read that here than find it out from a failed request.

Read this before you plug anything in. Inference is offline — no GPU is attached to the serving side, so /v1/models and /v1/chat/completions return 503. A miner you connect today accrues credits it cannot yet spend. The ledger is real and independent of the serving backend, so nothing you earn is lost — but you are connecting to start accruing, not to start calling.

1. Which miner for which card

The two stratum endpoints speak different protocols and are not interchangeable.

2. Install the miner

We do not build, sign, host or mirror either miner. Both are third-party software and we deliberately link you at the vendors rather than at a copy we control. Run the checksum step — alpha-miner is a closed binary from a third-party GitHub account and we have no way to vouch for its contents. If you would rather not run an unauditable binary on a machine that matters, put it on a box you can afford to lose.

alpha-miner, Linux — straight from the vendor's own README (releases, AlphaMine-Tech):

curl -L -o alpha-miner https://github.com/AlphaMine-Tech/alpha-miner/releases/latest/download/alpha-miner
chmod +x alpha-miner
curl -L https://github.com/AlphaMine-Tech/alpha-miner/releases/latest/download/SHA256SUMS | sha256sum -c

SRBMiner-MULTI is doktor83's multi-algo miner; pearlhash support landed upstream in 3.4.7. Download it from its own GitHub releases and verify the vendor's published checksums yourself.

3. Run it

Replace prl1pYOUR_ADDRESS_HERE with the Taproot PRL address you registered in Get your API key. That address is what shares accrue to, so use one whose private key you actually control.

alpha-miner on pearl/v1, port 5570. This is the path proven on real hardware — RTX 3090, alpha-miner 1.8.3, 139 of 139 shares accepted.

./alpha-miner --pool stratum+tcp://pool.linkrra.com:5570 --address prl1pYOUR_ADDRESS_HERE --worker rig1 --password 'x;d=64'

The flags are --pool, --address, --worker, --password. There is no --user and no --pass; alpha-miner aborts with unknown argument on anything it does not recognise. Keep it on one line — backslash continuations get mangled by some terminals and the miner then dies with unknown argument: " ".

SRBMiner-MULTI on pearlHash, port 5567. Use your full PRL address as the wallet; the pool derives your worker name from it.

SRBMiner-MULTI --algorithm pearlhash --pool pool.linkrra.com:5567 --wallet prl1pYOUR_ADDRESS_HERE --password x

4. Confirm it is working

A healthy first minute looks like this: the miner connects once and stays connected, it receives work, and the log starts printing accepted shares. Repeated reconnects with no accepted share is the failure described above, not a slow start.

You do not have to take our word that the endpoints are alive. Both answer immediately on connect — 5570 speaks first with a challenge, 5567 acknowledges a subscribe:

nc pool.linkrra.com 5570
# {"id":null,"method":"pearl.challenge","params":{"seed":"...","difficulty":8}}

printf '{"id":1,"method":"mining.subscribe","params":["probe"]}\n' | nc pool.linkrra.com 5567
# {"id":1,"result":[[["mining.set_difficulty","conn-..."],["mining.notify","conn-..."]],"...",4],"error":null}

The conn-... id differs on every connection; the shape is what matters. If you get neither of those, you are not talking to the pool.

Then check what your address has earned. Your sk-pool-... key is a standard bearer token:

curl -s https://ai.linkrra.com/v1/credits/balance -H "Authorization: Bearer sk-pool-YOUR_KEY"

{"address":"prl1p...","balance":0,"totalEarned":0,"totalSpent":0,
 "currency":"credits (1 credit = 1 token)"}

A wrong or unknown key gets you HTTP 401 with {"error":{"message":"Invalid or revoked API key."}}. Allocation runs on a schedule, so a freshly connected miner reads zero for a while before the first allocation lands — that is normal and not a sign your shares were dropped. Usage history is at /v1/credits/usage with the same header.

You need a balance of at least 100 credits before the API will accept a completion request; below that it answers HTTP 402 with insufficient_credits. That threshold is moot today — inference is offline regardless, and every completion returns 503 until GPU capacity is attached.

When it does not work

Still stuck? weare@linkrra.com. Live pool stats and fees: pool.linkrra.com.

Endpoints

The two protocols are not interchangeable. Aimed at the wrong one, a miner reconnect-loops silently instead of reporting an error.

Both miners are third-party software we do not build, sign or host — alpha-miner is distributed by AlphaMine-Tech as a closed binary, SRBMiner-MULTI by doktor83. Download them from the vendors' own GitHub releases above and verify the published SHA256SUMS yourself. We link the source, not a copy.

Endpoints re-verified on 2026-08-02, after DNS moved to the pool host. Until that day this name was a CDN CNAME that accepted the connection and then never answered, so older guides tell you to use the bare address 136.244.67.104 — that still works and is the fallback if name resolution is ever in doubt. Full mining setup, fees and live pool stats: pool.linkrra.com.