OpenAI-compatible · open-weight · verified mesh
The world's idle computers,
serving tokens.
Teraflock serves open-weight models from a verified mesh of everyday machines — Apple Silicon, gaming GPUs — through the API you already use. For work that doesn't need a frontier model, it's 5–20× cheaper. For the machines, it's a paycheck.
from openai import OpenAI client = OpenAI( base_url="https://api.teraflock.ai/v1", # the only change api_key="tf-...", ) r = client.chat.completions.create( model="flock/llama-3.1-8b-instruct", messages=[{"role": "user", "content": "hi"}], )
For builders
Drop-in API. Honest prices.
Every endpoint you expect
Chat completions with streaming, completions, embeddings — any OpenAI SDK works unmodified. Your framework doesn't know the difference.
Batch is first-class
Bulk classification, extraction, dataset generation, eval runs:
upload a file to /v1/batches and pay 0.7×
the interactive rate. Where latency doesn't matter, price should.
Verified, not trusted
Canary sampling against reference hardware, model fingerprinting, timing envelopes, and escrowed earnings make cheating economically irrational — so you don't have to take anyone's word for it.
Privacy, stated plainly
You choose where your prompts may travel.
Inference needs plaintext on the serving machine — anyone who says otherwise is selling something. So the tier is explicit, per request.
Any qualified node. Cheapest. Operators could theoretically observe content — we say so instead of pretending.
Only high-reputation nodes with staked credits and a signed operator agreement.
First-party datacenter hardware only. Highest on the mesh, still under frontier pricing.
On every tier: nodes see an ephemeral request ID, never your identity, and never write prompts to disk.
Pricing
Per million tokens. No subscriptions, no minimums.
| Class | Interactive | Batch (0.7×) | |
|---|---|---|---|
| Nano · ≤3B | $0.04 | $0.028 | classification, extraction, routing |
| Small · 7–9B | $0.1 | $0.07 | the volume workhorse |
| Mid · 13–34B | $0.3 | $0.21 | premium consumer hardware |
| Large · 70B | $0.7 | $0.49 | scarce supply, priced accordingly |
| Embeddings | $0.01 | $0.007 | bulk vectorization |
Full pricing → including reputation and surge multipliers.
For operators
Your computer sleeps eight hours a night.
It could be earning.
Two minutes to enroll
One daemon. It detects your hardware, pulls the right models, and starts serving. Mac menu bar, CLI, and a local dashboard.
Never in your way
Serves only when you're away by default; touch the keyboard and it yields within two seconds. Battery, hours, and memory limits are yours.
Real money, honestly framed
Operators keep the majority of what customers pay. A gaming PC won't fund your retirement — it can pay its own electricity and your AI bill. Big memory earns premium rates.
FAQ
Fair questions, straight answers.
Is it really OpenAI-compatible?
Yes — chat completions, completions, embeddings, and batches. Point any OpenAI SDK at our base URL, keep your code. Streaming included.
Where do my prompts actually go?
To someone's computer — and we say so. Requests carry an explicit privacy tier: open routes to any qualified node (cheapest; operators could theoretically observe content), verified routes only to high-reputation, staked, under-agreement nodes, and private stays on first-party hardware. Nodes never see who you are, and never persist prompts to disk.
How do you know a node isn't faking the work?
Statistically and economically. A slice of requests is silently duplicated to reference hardware and compared; nodes get fingerprint challenges with known answers for their exact model build; timing outside the hardware's physical envelope gets flagged. Cheaters forfeit escrowed earnings and their reputation — cheating has negative expected value.
What can my machine earn?
Honest answer: it depends on demand. A typical gaming PC serving an 8B model a few idle hours a day should think in tens of dollars a month at launch, not mining-boom numbers — the frame is a GPU that pays its own electricity and funds your own AI usage. Bigger memory earns premium rates: 24GB+ VRAM or 64GB+ unified memory serves the models where supply is scarce.
Will it slow my computer down?
That promise is the whole product. By default the daemon serves only when you're away, and the moment you touch the keyboard it yields within two seconds — the mesh reroutes without you noticing. You set the limits: hours, battery rules, memory caps.
Why is batch cheaper?
Because that's where a mesh structurally wins. Embeddings, bulk classification, dataset generation, eval runs — work where price matters and latency doesn't — run at 0.7× the interactive rate through a first-class /v1/batches API.
When can I use it?
We're onboarding from the waitlist in cohorts — builders and operators in matched batches, because the marketplace needs both sides. Join below; we'll write when it's your turn.
Get in the flock.
Builders and operators onboard in matched cohorts — the marketplace needs both sides. Tell us which you are: